Movie dubbing inference methods, devices, equipment, media and systems that support multimodal parallel processing

By generating modal track-based load records and performing time-series rewriting calibration, the problems of low processing path utilization and repeated boundary adjustments in existing translation schemes are solved. This enables parallel processing of multi-source content and continuous timeline rewriting, improving translation efficiency and consistency.

CN123075145APending Publication Date: 2026-11-10CHINA FILM IND GROUP CO LTD BEIJING ARTIFICIAL INTELLIGENCE RESEARCH & APPLICATION BRANCH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611019523.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-09
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing film translation processing solutions struggle to balance processing efficiency and write-back consistency, especially in feature films, multi-character, and multi-scene switching tasks. They cannot effectively accommodate cross-segment relationships and parallel execution between different data contents, leading to issues such as repeated adjustments to boundary connections and low path utilization in the processing results.

Method used

By generating modal track-based load records, parallel load channels are allocated to translated inference segments using segment anchors and modal dependency identifiers. Channel scheduling records are generated, and channel translation results are generated within the reloadable time period defined by the segment anchors. Finally, when the retrieval readiness conditions are met, timing reload calibration is performed to ensure the consistency of the results with the movie timeline.

Benefits of technology

It improves the utilization rate of the processing path in the translation and reasoning process, reduces the risk of premature processing of dependent fragments caused by disordered parallelism, reduces write-back waiting and boundary conflicts between parallel results, and realizes the requirements of parallel processing of multi-source content and continuous write-back of the time axis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123075145A_ABST
    Figure CN123075145A_ABST
Patent Text Reader

Abstract

This application provides a method, apparatus, electronic device, computer-readable storage medium, and system for movie translation inference that supports multimodal parallel processing. The method acquires movie data to be translated and target language translation configuration, generates modal track-based carrying records according to the presentation time sequence of heterogeneous modal content, extracts translation inference segments and allocates parallel carrying channels, drives each channel to generate channel translation results and writes them to the recycling cache, and performs time-series backflow calibration when the recycling ready conditions are met to obtain the movie translation inference results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of film translation and multimodal data processing technology, and more specifically, to a film translation inference method, apparatus, electronic device, computer-readable storage medium and system that supports multimodal parallel processing. Background Technology

[0002] With the increasing demand for cross-language dissemination of film and television content, the film dubbing process no longer focuses solely on the conversion of target language text. It also needs to simultaneously consider the rhythm of the original audio, changes in visual action, semantic continuity of subtitles, and the audiovisual continuity of the dubbed content after being written back to the film's timeline. Especially in the dubbing of feature films with multiple characters and multiple scene transitions, film data typically includes data from different sources such as audio, visuals, and text. These various data contents have different granularities of presentation on the timeline, requiring the dubbing reasoning process to achieve a good balance between processing efficiency and consistency in the write-back.

[0003] Existing film dubbing solutions typically employ a centralized, sequential processing workflow. This approach first splits the film file, extracts the audio track, subtitles, or dialogue text, and generates the target language content segment by segment in chronological order. Subsequently, each segment of target language content undergoes speech synthesis or subtitle conversion, and is written back to the film file according to its original time position. Finally, the translated file is adjusted through an overall preview or unified verification method.

[0004] However, the aforementioned sequential processing scheme has significant technical flaws. Since data from different sources is typically processed in a uniform linear order, it's difficult to simultaneously consider cross-segment relationships between different data contents and the parallel execution of various processing tasks during the inference process. This can easily lead to situations where subsequent segments need to read information from preceding segments while the processing flow has already progressed. Furthermore, the sequential processing scheme lacks process constraints oriented towards the result write-back boundary. When writing back multiple segments of translated content in a concentrated manner, it can easily result in waiting for processing results, repeated adjustments to boundary connections, and low utilization of processing pathways, making it difficult to meet the processing requirements of parallel inference of multi-source content and continuous timeline write-back in film translation. Summary of the Invention

[0005] This application provides a movie translation reasoning method, apparatus, electronic device, computer-readable storage medium, and system that support multimodal parallel processing, so as to at least alleviate the above-mentioned technical problems.

[0006] A movie dubbing inference method supporting multimodal parallel processing includes: The system acquires the film data to be translated and the target language translation configuration. Based on the presentation sequence of various heterogeneous modal content in the film data to be translated, it generates a modal track carrying record. The modal track carrying record includes segment anchors and modal dependency identifiers. The segment anchors are used to define the reloadable time periods of the translated content on the film timeline. The modal dependency identifiers are used to represent the cross-segment reading relationship between different segment anchors. Based on the segment anchor point, the translated inference segment is extracted from the modal track-based bearer record, and a parallel bearer channel is allocated to each translated inference segment based on the modal dependency identifier, generating a channel scheduling record; the channel scheduling record records the segment anchor point and its reading status corresponding to each parallel bearer channel; Based on the channel scheduling record, each of the parallel bearer channels is driven to read the corresponding translation inference fragment, so that each parallel bearer channel generates the channel translation result within the recycle period defined by the fragment anchor point, and writes the channel translation result into the recycling cache indexed by the fragment anchor point; When the channel translation results belonging to the same segment anchor point in the recycling cache are detected to meet the recycling readiness conditions, the channel translation results are subjected to time-series backflow calibration to obtain the movie translation inference results.

[0007] Optionally, generating modal track carrying records based on the presentation sequence of multiple heterogeneous modal content in the film data to be translated includes: Parse the encapsulation index in the film data to be translated, and establish a unified timeline based on the timestamp information in the encapsulation index; Based on the presentation sequence of the heterogeneous modal content, the original sound segments are marked on the unified time axis; The text carrier segment is mapped to the segment anchor point corresponding to the original sound segment to obtain the text carrier mapping result, and the text carrier mapping result carries the segment anchor point corresponding to the text carrier segment; The on-screen action segment is mapped to the segment anchor point corresponding to the original sound segment to obtain the on-screen action mapping result, and the on-screen action mapping result carries the segment anchor point corresponding to the on-screen action segment. The modal track-based carrying record is generated based on the text carrying mapping result and the screen motion mapping result, and when there is a cross-segment continuation relationship, the modal dependency identifier is written to the modal track-based carrying record.

[0008] Optionally, the modal dependency identifier is generated based on the continuous state of the frame and the continuous state of the original sound at adjacent segment anchor points, including: Obtain the video action segments corresponding to adjacent segment anchor points from the video action mapping results, and determine the video continuity state based on the content continuity of the video action segments; Obtain the sound boundary corresponding to the anchor point of the adjacent segment from the original sound segment, and determine the continuous state of the original sound based on the sound boundary; When both the continuous state of the video and the continuous state of the original sound indicate the existence of a cross-segment continuation relationship, a modal dependency identifier pointing to the anchor point of the preceding segment is generated, and the modal dependency identifier is written into the modal track carrying record corresponding to the current segment anchor point.

[0009] Optionally, the step of extracting the translated reasoning segment from the modal track-based record based on the segment anchor point includes: Read the segment anchor point and the modal dependency identifier from the modal track bearing record; Based on the segment anchor point, extract the source segment content corresponding to the segment anchor point from the movie data to be translated; Obtain the target language text corresponding to the source fragment content from the target language translation configuration; The sound duration constraint is generated based on the rechargeable time period defined by the segment anchor point; Image fitting constraints are generated based on the image motion mapping results in the modal track-splitting record. Based on the source fragment content, the target language text, the speech duration constraint, the image fitting constraint, and the modal dependency identifier, the translated reasoning fragment is generated, and the modal dependency identifier is used as the dependency carrier field of the translated reasoning fragment.

[0010] Optionally, the step of allocating parallel bearer channels for each translated inference segment based on the modal dependency identifier and generating channel scheduling records includes: Obtain a pre-configured list of parallel bearer channels, which is pre-established according to the target language translation configuration, and record the channel identifier and cache status of each parallel bearer channel; Extract the modal dependency identifier recorded in the dependency carrier field of the translated reasoning fragment; Based on the cache state and the modal dependency identifier, the parallelism of each translated inference segment is determined; Place the translation and inference segments that are in a parallelizable state into the waiting slot of the corresponding parallel bearer channel, and place the translation and inference segments that are not in a parallelizable state into the delay slot of the corresponding parallel bearer channel. The channel scheduling record is generated based on the segment anchor points and reading status corresponding to the slot to be processed and the delay slot, respectively.

[0011] Optionally, the step of driving each of the parallel bearer channels to read the corresponding decoded inference fragments according to the channel scheduling record includes: Based on the segment anchor points and reading status in the channel scheduling record, set the anchor point gatekeeper status for each parallel bearer channel; When the anchor gatekeeper status indicates that there are no unmet modal dependency identifiers in the translated inference fragment in the slot to be processed, the corresponding parallel bearer channel is allowed to read the translated inference fragment; When the anchor point of the preceding segment that the translated inference segment in the delay slot depends on is marked with a channel completion identifier in the channel scheduling record, the gatekeeper status of the anchor point is updated according to the channel completion identifier, and the translated inference segment is transferred from the delay slot to the waiting slot for the corresponding parallel bearer channel to read.

[0012] Optionally, the channel translation result includes source segment anchors, channel completion identifiers, re-translated content, and boundary preservation tags; The process of generating channel decoding results for each parallel bearer channel within the rechargeable time period defined by the segment anchor point includes: Each parallel bearer channel reads the corresponding translated inference segment within the rechargeable time period defined by the segment anchor point; Based on the target language text, speech duration constraints, and image fitting constraints in the translated reasoning segment, generate re-translated content; Based on the rechargeable time period defined by the segment anchor point, a boundary preservation label containing start and end protection points is generated, wherein the start and end protection points are used to indicate the start and end positions of the rechargeable time period; Write the segment anchor point as a homologous segment anchor point into the channel translation result, and write the channel completion identifier into the channel translation result; The channel translation result carrying the same source segment anchor point is written into the recycling cache, and the channel completion identifier is written back to the channel scheduling record.

[0013] Optionally, detecting that the channel translation results belonging to the same segment anchor point in the reclaimed cache meet the reclaim ready condition includes: Taking the channel translation result corresponding to the same source segment anchor point in the reclaimed cache as the object, query the registered parallel bearer channel in the channel scheduling record according to the same source segment anchor point; When all the registered parallel bearer channels have a channel completion identifier, it is confirmed that the channel recycling condition is met; When the end protection point of the previous segment anchor point and the start protection point of the next segment anchor point are connected on the time axis among the start and end protection points indicated by the boundary preservation labels of adjacent segment anchor points, the boundary preservation label is confirmed to have passed the consistency check. When the re-translatable content meets the speech duration constraint and the image fit constraint, it is confirmed that the content re-translation condition is met. When the channel reclamation condition, the consistency check, and the content re-injection condition are all satisfied, it is determined that the reclamation readiness condition is satisfied. If any condition is not met, the corresponding channel translation result will be kept in the recycling cache, and the channel translation result will be prohibited from entering the timing backfill calibration.

[0014] Optionally, performing timing backfeed calibration on the channel decoding result includes: Taking the channel translation results that meet the recycling readiness conditions as the object, the corresponding rechargeable time period on the time axis is determined according to the homologous segment anchor point; The start and end protection points of the rechargeable period are determined based on the boundary preservation label; Within the range defined by the start and end protection points, the re-translatable content is written into the re-translatable time period; Based on the write-throughable period, a calibrated anchor state is generated and written back to the channel scheduling record so that the translation inference segment in the delay slot can be restored according to the calibrated anchor state to obtain the movie translation inference result.

[0015] This application also provides a movie dubbing inference device that supports multimodal parallel processing, including: The modal track-based delivery module is used to acquire the film data to be translated and the target language translation configuration, and generate modal track-based delivery records according to the presentation sequence of multiple heterogeneous modal content in the film data to be translated. The modal track-based delivery record includes segment anchors and modal dependency identifiers. The segment anchors are used to define the re-feedable time period of the translated content on the film timeline, and the modal dependency identifiers are used to represent the cross-segment reading relationship between different segment anchors. The channel scheduling record generation module is used to extract the translated inference segment from the modal track-based bearer record according to the segment anchor point, and allocate parallel bearer channels for each translated inference segment based on the modal dependency identifier, thereby generating a channel scheduling record; the channel scheduling record records the segment anchor point and its reading status corresponding to each parallel bearer channel; The parallel bearer execution module is used to drive each of the parallel bearer channels to read the corresponding translation inference fragments according to the channel scheduling record, so that each parallel bearer channel generates the channel translation result within the recycle period defined by the fragment anchor point, and writes the channel translation result into the recycling cache indexed by the fragment anchor point. The timing backfeed calibration module is used to perform timing backfeed calibration on the channel translation results when the channel translation results belonging to the same segment anchor point in the recycling cache meet the recycling ready conditions, so as to obtain the movie translation inference results.

[0016] This application also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the aforementioned movie translation inference method that supports multimodal parallel processing.

[0017] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned movie translation reasoning method supporting multimodal parallel processing.

[0018] This application also provides a movie dubbing inference system that supports multimodal parallel processing, including a movie data access node, a channel scheduling node, a parallel carrying node, and a recycling and recharge node; The film data access node is used to acquire the film data to be translated and the target language translation configuration. Based on the presentation sequence of multiple heterogeneous modal content in the film data to be translated, it generates a modal track-based carrying record and sends the modal track-based carrying record to the channel scheduling node. The modal track-based carrying record includes segment anchors and modal dependency identifiers. The segment anchors are used to define the re-feedable time period of the translated content on the film timeline, and the modal dependency identifiers are used to represent the cross-segment reading relationship between different segment anchors. The channel scheduling node is used to extract the translated inference segment from the modal track-based bearer record according to the segment anchor point, allocate parallel bearer channels for each translated inference segment based on the modal dependency identifier, generate a channel scheduling record, and send the channel scheduling record to the parallel bearer node; the channel scheduling record records the segment anchor point and its reading status corresponding to each parallel bearer channel; The parallel bearer node is used to drive each of the parallel bearer channels to read the corresponding translation inference fragments according to the channel scheduling record, so that each parallel bearer channel generates the channel translation result within the recycle period defined by the fragment anchor point, and writes the channel translation result into the recycling cache indexed by the fragment anchor point. The recycling and recharge node is used to perform time-series recharge calibration on the channel translation results when the channel translation results belonging to the same segment anchor point in the recycling cache meet the recycling readiness conditions, so as to obtain the movie translation inference results.

[0019] This application also provides a movie dubbing inference method that supports multimodal parallel processing, including: Obtain the movie data to be processed and the translation constraint configuration, and generate time-carrying information based on the multimodal temporal correlation in the movie data to be processed. The time-carrying information is used to limit the processing boundary of the translated reasoning content in the movie time sequence and the reading dependency relationship between different processing boundaries. A translation inference bearer object is constructed based on the timing bearer information, and a parallel execution guide is configured for the translation inference bearer object according to the reading dependency relationship. The parallel execution guide is used to indicate the bearer position and reading status of each translation inference bearer object. According to the parallel execution guidelines, multiple parallel processing paths are driven to read the corresponding translation inference bearer objects, so that the multiple parallel processing paths generate path translation results under the processing boundary constraints, and write the path translation results into the result recycling area; When the path translation results in the result recovery area meet the recovery judgment conditions used to characterize the recirculation of multiple path results within the same processing boundary, the path translation results are subjected to time-series calibration processing to obtain the movie translation inference results.

[0020] The technical advantages of the technical solution provided in this application are: This application's movie translation inference method, which supports multimodal parallel processing, addresses the technical shortcomings of existing sequential processing schemes that struggle to simultaneously consider cross-segment relationships and parallel execution relationships between different data contents. It first acquires the movie data to be translated and the target language translation configuration, and then generates modal track-based carrying records based on the presentation sequence of various heterogeneous modal content in the movie data. This allows content from different sources in the movie to be converted into a record format that can be carried according to the temporal presentation relationship. Simultaneously, it defines the re-feedback periods of the translated content on the movie timeline through segment anchors, and represents the cross-segment reading relationships between different segment anchors through modal dependency identifiers. Compared to traditional schemes that process only segment by segment according to file time sequence, this application can simultaneously establish time boundaries and cross-segment reading relationships before translation inference begins. This ensures that subsequent parallel processing does not deviate from the write-back constraints on the movie timeline, nor ignore the reading dependencies between adjacent or related segments, thereby alleviating the problem in sequential processes where subsequent segments need to read information from preceding segments but lack carrying identifiers.

[0021] Based on the existing modal track-based load tracing record, this application extracts translated inference segments from the record according to segment anchor points, and allocates parallel load tracing channels to each translated inference segment based on modal dependency identifiers. It also generates a channel scheduling record that records the segment anchor points and their reading states for each parallel load tracing channel. Compared to traditional methods that process all segments sequentially in a single sequential queue, this application does not simply split the movie content into segments and execute them concurrently. Instead, it uses segment anchor points to determine the processing objects of the translated inference segments, and then uses modal dependency identifiers to determine the allocation relationship between the translated inference segments and the parallel load tracing channels. This allows segments that can be read in parallel to enter the parallel processing process, while segments with cross-segment reading relationships are constrained by their reading states. Therefore, this approach improves the utilization rate of translated inference segments across different channels, reduces the risk of premature processing of dependent segments due to disordered parallelism, and solves the problems of low processing path utilization and the potential disruption of segment relationships after parallel modification in existing sequential processing schemes.

[0022] This application uses channel scheduling records to drive each parallel carrier channel to read the corresponding translation and inference segments, enabling each parallel carrier channel to generate channel translation results within the recycleable time period defined by the segment anchor point, and write the channel translation results into a recycling cache indexed by the segment anchor point. Through this processing method, each channel translation result is associated with its corresponding segment anchor point and recycleable time period during generation. The channel translation results are not directly written back to the movie timeline after generation, but first enter the recycling cache indexed by the segment anchor point. Compared to the traditional approach where each segment of translated content is directly written back to its original position after generation and then uniformly checked later, this application separates parallel generation and time-series write-back, allowing parallel carrier channels to generate results at their respective processing rhythms. Simultaneously, the results are aggregated to the corresponding recycling positions through the segment anchor point index, thereby reducing write-back waiting and boundary conflicts caused by different generation sequences between parallel results. This provides a data foundation for subsequent unified determination of whether results under the same segment anchor point are recycleable.

[0023] When the translated channel results belonging to the same segment anchor point in the recycling cache are detected to meet the recycling readiness condition, this application performs time-series backfeed calibration on the channel translated results to obtain the movie translation inference results. This processing method addresses the problem that existing sequential processing schemes lack process constraints oriented towards the result write-back boundaries and are prone to repeated adjustments of boundary connections during concentrated write-back. By adding a recycling readiness judgment before the channel translated results enter the movie timeline, the translated channel results belonging to the same segment anchor point only enter the time-series backfeed calibration when they meet the backfeedability requirements. Compared to the traditional approach of writing back first and then previewing and adjusting the whole, this application uses the recycling cache and recycling readiness conditions as pre-constraints for parallel results to enter the timeline, and writes the backfeedable translated content into the movie timeline according to the time boundaries corresponding to the segment anchor points through time-series backfeed calibration. This results in better consistency between the multi-channel generated results and the movie timeline, reduces the amount of repeated adjustments caused by concentrated write-back of multiple translated content, and makes the movie translation inference process more suitable for the application requirements of parallel processing of multi-source content and continuous backfeed. Attached Figure Description

[0024] Figure 1 This application provides a scenario for a movie dubbing inference method that supports multimodal parallel processing. Figure 2 This application provides an embodiment of a movie translation inference method that supports multimodal parallel processing. Figure 3 This application provides an embodiment of a movie dubbing inference device that supports multimodal parallel processing. Figure 4 An electronic device is described in an embodiment of this application; Figure 5 This application provides an embodiment of a computer-readable storage medium. Figure 6 This application provides an embodiment of a movie dubbing inference system that supports multimodal parallel processing. Figure 7 This application provides an embodiment of a movie translation inference method that supports multimodal parallel processing. Detailed Implementation

[0025] like Figure 1 The image shown illustrates a scenario of a movie translation inference method supporting multimodal parallel processing, according to an embodiment of this application. Figure 2As shown in the figure, this application provides an embodiment of a film translation inference method that supports multimodal parallel processing. This method addresses scenarios where multiple heterogeneous modalities, such as original audio data, original text data, and on-screen action segments, exist simultaneously on the same film timeline during film translation. Film translation is not simply a conversion of a text into a target language; rather, it involves placing the translated target language content back into a position on the film timeline where re-feedback is permitted. Furthermore, it is crucial that the continuity relationships between adjacent segments—including the continuity of original audio segments, the continuity of on-screen action segments, and the correspondence between text segments—are not disrupted during parallel processing. To this end, before translation inference, this application first converts the various heterogeneous modalities scattered across different data tracks in the film data to be translated into modal track-based transport records that can be uniformly scheduled. Then, using the segment anchor points and modal dependency identifiers in the modal track-based transport records, parallel-processable translation inference segments are allocated to different parallel transport channels. Translation inference segments with cross-segment reading relationships retain the necessary reading order, thereby establishing a data structure correspondence between the parallel processing process and the re-feedbackable time periods on the film timeline.

[0026] Specifically, such as Figure 2 As shown, the movie translation inference method supporting multimodal parallel processing includes the following steps: Acquire the film data to be translated and the target language translation configuration, and generate a modal track carrying record based on the presentation sequence of multiple heterogeneous modal content in the film data to be translated; Based on the segment anchor point, the translated inference segment is extracted from the modal track-based bearer record, and parallel bearer channels are allocated to each translated inference segment based on the modal dependency identifier, generating a channel scheduling record; Based on the channel scheduling record, each parallel bearer channel is driven to read the corresponding translation inference fragment, so that each parallel bearer channel generates the channel translation result within the recycle period defined by the fragment anchor point, and writes the channel translation result into the recycling cache indexed by the fragment anchor point. When the channel translation results belonging to the same segment anchor point in the recycling cache are detected to meet the recycling readiness conditions, the channel translation results are subjected to time-series backflow calibration to obtain the movie translation inference results.

[0027] The process involves acquiring the film data to be translated and the target language translation configuration, and generating a modal track carrying record based on the presentation sequence of various heterogeneous modal content in the film data. In this application, the film data to be translated can be a film file containing image data, original audio data, original text data, and a container index, or it can be a film data package in which image data, original audio data, original text data, and a container index are stored separately but share the same film timeline. The film data to be translated is not limited to a specific container format; as long as it contains a container index describing the positions of image data, original audio data, and original text data on the film timeline, it can be used as input for this application. The target language translation configuration is the translation processing configuration corresponding to the film data to be translated. The target language translation configuration can be read from the translation task configuration file, translation task request data, or a pre-written configuration table. The translation task configuration file, translation task request data, and pre-written configuration table all serve as sources for reading the target language translation configuration. The target language translation configuration at least indicates the source of the target language text, the target language output format, the available parallel carrying channels, and the constraints on the location of the translated content feedback. After the film data to be translated and the target language translation configuration are acquired, a unified timeline is first established according to the encapsulation index in the film data to be translated, enabling the image data, original audio data, and original text data to be read on the same film timeline. The presentation sequence of multiple heterogeneous modal content refers to the order of appearance, duration, and adjacency of image data, original audio data, and original text data on the film timeline. The presentation sequence is not the reading order of a single file, but rather the temporal relationship obtained after placing the image data, original audio data, and original text data onto a unified timeline. The modal track-based carrier record generated according to the presentation time sequence is used to carry segment information of different data tracks on the movie timeline. Each record in the modal track-based carrier record is associated with a segment anchor point, and the segment anchor point indicates the reloadable period of the record on the movie timeline.

[0028] In the above processing, segment anchors are used to unify the generation and re-feeding positions of the translated content onto the movie timeline. A segment anchor can consist of a segment start time marker, a segment end time marker, and a segment anchor identifier. The segment start time marker indicates the position on the movie timeline where the translated content is allowed to begin re-feeding, the segment end time marker indicates the position on the movie timeline where the translated content is allowed to end re-feeding, and the segment anchor identifier is used to maintain index consistency among subsequent translated inference segments, channel scheduling records, channel translation results, and cache recycling. Modal dependency identifiers are used to represent cross-segment reading relationships between different segment anchors. Modal dependency identifiers can be written into the modal track-carrying record corresponding to the current segment anchor and point to the preceding segment anchor that needs to be read. If the current segment anchor needs to read the sound boundaries, visual action segments, or target language text connection information of the preceding segment anchor before generating the translated content, the modal track-carrying record corresponding to the current segment anchor will carry the modal dependency identifier. Therefore, the modal track-based carrying record not only saves the time position of the current segment anchor point on the movie timeline, but also saves the reading constraints between the current segment anchor point and the previous segment anchor point, so that subsequent parallel carrying channels can determine whether the translated inference segment can be directly entered into parallel processing based on the modal dependency identifier.

[0029] Based on the segment anchor point, translated inference segments are extracted from the modal track-based transport record, and parallel transport channels are allocated to each translated inference segment based on the modal dependency identifier, generating a channel scheduling record. In this application, the translated inference segment is a processable data unit extracted from the modal track-based transport record according to the segment anchor point. The translated inference segment is not only composed of text content, but also carries the source segment content corresponding to the segment anchor point, the target language text, the sound duration constraint, the picture fitting constraint, and the modal dependency identifier. The source segment content is used to provide the picture segment, original audio segment, and original text segment corresponding to the segment anchor point in the movie data to be translated. The target language text is used to provide the target language translated content that needs to be generated by the current segment anchor point. The sound duration constraint is used to limit the time occupation of the target language translated content within the reloadable period. The picture fitting constraint is used to limit the correspondence between the target language translated content and the picture action segment. The modal dependency identifier is used to indicate whether the translated inference segment needs to wait for the processing status of the previous segment anchor point before entering the parallel transport channel. Parallel bearer channels are bearer locations used for parallel reading of translated inference fragments and generation of channel translation results. Parallel bearer channels can correspond to different computation threads, different task queues, or different parallel execution processes, but in this application, they all participate in the generation of channel scheduling records in the form of parallel bearer channels. Channel scheduling records are used to record the fragment anchor points and their reading states corresponding to each parallel bearer channel. The reading state can indicate that the translated inference fragment is in a pending slot, a delayed slot, being read, or has been completed. The reading state will be read by subsequent parallel bearer channels and will affect whether the translated inference fragment is allowed to be processed in parallel.

[0030] Based on the channel scheduling record, each parallel bearer channel is driven to read its corresponding translation inference segment. This allows each parallel bearer channel to generate its translation result within the recycleable time period defined by the segment anchor point, and write the translation result to a recycling cache indexed by the segment anchor point. In this application, the channel scheduling record serves as the basis for reading by each parallel bearer channel after its generation. The parallel bearer channel does not directly traverse the movie data to be translated; instead, it reads the translation inference segment already allocated to it based on the segment anchor point, reading status, and parallel bearer channel correspondence recorded in the channel scheduling record. After reading the translation inference segment, the parallel bearer channel needs to generate its translation result within the recycleable time period defined by the segment anchor point. Therefore, the translation result carries the recycleable time period corresponding to the segment anchor point during generation, rather than being temporarily matched to the movie timeline after generation. The recycling cache is a cache area used to temporarily store the translation results of the channels. The recycling cache uses the segment anchor point as an index, allowing translation results belonging to the same segment anchor point to be read uniformly subsequently. By writing the channel translation results into the recycling cache first, the channel translation results can still be aggregated to the corresponding segment anchor point at different completion times, avoiding the problem of the channel translation results directly entering the movie timeline and causing difficulty in coordinating the reflow order and segment boundaries.

[0031] When the channel translation results belonging to the same segment anchor point in the recycling cache are detected to meet the recycling readiness conditions, a temporal backfeed calibration is performed on the channel translation results to obtain the movie translation inference results. In this application, the recycling readiness conditions are used to determine whether the channel translation results corresponding to the same segment anchor point have met the conditions for entering the movie timeline. The recycling readiness conditions may include that all registered parallel carrier channels under the same segment anchor point have channel completion identifiers, the channel translation results meet the sound duration constraints, the channel translation results meet the screen fitting constraints, and the boundary preservation labels between adjacent segment anchor points can be continued on the movie timeline. The channel translation results will not be written to the movie timeline before meeting the recycling readiness conditions. The temporal backfeed calibration is used to write the re-feedable translation content in the channel translation results to the movie timeline according to the re-feedable time period defined by the segment anchor point, and form the calibrated anchor point state according to the re-feedable time period after writing. The movie translation inference results are translation result data formed after temporal backfeed calibration, which at least includes the backfeed position of the target language translation content on the movie timeline and the segment anchor point corresponding to the backfeed position. The above processing ensures that the channel translation results generated by the parallel carrier channel are first collected through the segment anchor point index, and then enter the time-series backflow calibration after being subject to the retrieval ready condition constraints, so that the parallel processing results can be consistent with the backflow boundary of the movie timeline.

[0032] Based on the presentation sequence of various heterogeneous modal content in the film data to be translated, a modal track carrying record is generated, including: This application parses the encapsulation index in the film data to be translated and establishes a unified timeline based on the timestamp information in the encapsulation index. In this application, the encapsulation index is the index information used to locate video data, original audio data, and original text data in the film data to be translated. The encapsulation index can include the storage location of each data packet in the file, the timestamp information corresponding to the data packet, and the data track identifier to which the data packet belongs. The timestamp information is a time marker in the encapsulation index used to indicate the position of the data packet in the film playback order. The timestamp information in different data tracks can use different timing starting points or different time granularities; therefore, the timestamp information in different data tracks cannot be directly used as a unified processing basis. When establishing a unified timeline, the timestamp information and data track identifier in the encapsulation index are first read. Then, the data tracks to which the video data, original audio data, and original text data belong are distinguished based on the data track identifier. Subsequently, the timestamp information belonging to the video data, original audio data, and original text data is converted to the same film timeline. The conversion process can use the film playback start point as the starting point of the unified timeline, and convert the timestamp information in different data tracks into time positions on the unified timeline based on the track start offset and playback order marker of each data track in the encapsulation index. Once a unified timeline is established, video data, original audio data, and original text data can all be read and linked together through the unified timeline. Subsequent original audio segments, text segments, and video action segments all use the unified timeline as the calibration basis.

[0033] Based on the presentation sequence of heterogeneous modal content, original audio segments are marked on a unified timeline. In this application, an original audio segment is the corresponding sound interval of the original audio data in the film data to be translated on a unified timeline. Original audio segments can be obtained by reading the start and end markers in the original audio data, or by performing speech activity recognition on the audio frames in the original audio data. To enable the original audio segments to participate in the generation of subsequent segment anchors, the original audio data is first divided into a continuous sequence of audio frames according to a unified timeline. Then, the start and end positions of the sound are determined based on the energy changes, silence intervals, and sound continuity in the audio frame sequence. The start and end positions together constitute the original audio segment. The original audio segment is not the final translated content, but rather the temporal basis for segment anchor generation and text-carrying segment mapping. Once the original audio segment is identified, the subsequent text segments will be mapped to the corresponding segment anchor point of the original audio segment, and the visual action segments will also be mapped to the segment anchor point of the same original audio segment, so that the text segments, visual action segments, and original audio segments have the same segment anchor point relationship on the same timeline.

[0034] The text carrier segment is mapped to the corresponding segment anchor point of the original audio segment to obtain the text carrier mapping result. The text carrier mapping result carries the segment anchor point corresponding to the text carrier segment. In this application, the text carrier segment is a text range in the film data to be translated that can provide the source of the translated text. The text carrier segment can come from the subtitle track, dialogue script text, or original text identified from the original audio data. Before entering the mapping process, the text occurrence position of the text carrier segment is first determined according to a unified timeline. If the text carrier segment itself has timestamp information, the timestamp information of the text carrier segment is converted to the unified timeline. If the text carrier segment does not have complete timestamp information, a correspondence is established based on the text order and sound order between the text carrier segment and the original audio segment. After determining the text occurrence position of the text carrier segment on the unified timeline, the text carrier segment is mapped to the segment anchor point corresponding to the original audio segment whose text occurrence position overlaps with or is adjacent to its text occurrence position to obtain the text carrier mapping result. The text carrier mapping result records at least the content storage location of the text carrier segment, the text occurrence position of the text carrier segment on the unified timeline, and the segment anchor point corresponding to the text carrier segment. The text-carrying mapping results serve as the carrier basis for the text source when generating modal track-segmented carrier records, and provide a location reference for the target language text when extracting translated reasoning fragments.

[0035] The process involves mapping on-screen action segments to corresponding segment anchors in the original audio segments to obtain the on-screen action mapping result. This result carries the corresponding segment anchors. In this application, the on-screen action segment is a segment of the film data to be translated that is related to the re-injection of the translated content. The on-screen action segment can include character voice-over scenes, lip-syncing scenes, action sequences before and after shot transitions, and scene segments related to the rhythm of the voice-over. Before entering the mapping process, the on-screen action segment's position on a unified timeline is determined based on the frame timestamps in the on-screen data. Then, the on-screen action segment is determined based on the frame transition markers and the range of motion changes between frames. The frame transition markers can be read from the shot segmentation information in the film data to be translated, or generated through the differences between adjacent frames. The range of motion changes can be determined by the opening and closing of the voice-over area in the frame, changes in facial orientation, or changes in motion synchronized with the voice-over in the frame. After determining the on-screen action segment, it is mapped to the segment anchor corresponding to the original audio segment that overlaps with its on-screen position to obtain the on-screen action mapping result. The image-action mapping result carries at least the image location of the image-action segment on a unified timeline, the segment anchor point corresponding to the image-action segment, and the image boundary information of the image-action segment. The image-action mapping result is read when generating image-fitting constraints, so that the target language text in the translated inference segment can maintain a time correspondence with the image-action segment within the refeedable time period defined by the segment anchor point.

[0036] Modal track-based carrier records are generated based on the text-carrying mapping results and the image-action mapping results. When cross-segment continuation relationships exist, modal dependency identifiers are written to the modal track-based carrier records. In this application, both the text-carrying mapping results and the image-action mapping results carry segment anchors, thus allowing them to be grouped into the same modal track-based carrier record according to their segment anchors. When generating modal track-based carrier records, the text-carrying mapping results are first read using the segment anchors as indexes, then the image-action mapping results with the same segment anchors are read, and finally, the text-carrying mapping results, image-action mapping results, and corresponding original sound segments are written into the same modal track-based carrier record. After the modal track-based carrier records are formed, each segment anchor can correspond to an original sound segment, a text-carrying segment, and an image-action segment, eliminating the need to search through all different data tracks in the film data to be translated when subsequently extracting translated inference segments. Cross-segment continuation relationships are used to indicate situations where there is a need to read preceding information between the current segment anchor and the previous segment anchor. When cross-segment continuation relationships exist, a modal dependency identifier is written to the modal track-based carrier record corresponding to the current segment anchor point. This modal dependency identifier points to the previous segment anchor point and is then entered into the dependency carrier field when generating subsequent decoded inference segments. In this way, the cross-segment continuation relationship is solidified into a readable data field at the modal track-based carrier record stage. Subsequent parallel carrier channels do not need to temporarily determine the relationship between adjacent segments but directly determine the reading order based on the modal dependency identifier.

[0037] Modal dependency identifiers are generated based on the continuity of the visuals and the continuity of the original audio at adjacent segment anchor points, including: The image action segments corresponding to adjacent segment anchors are obtained from the image action mapping results. The image continuity state is determined based on the content continuity of the image action segments. In this application, the image continuity state is used to indicate whether there is a cross-segment continuation relationship between image action segments between adjacent segment anchors. When determining the image continuity state, the image action segment corresponding to the previous segment anchor is first read from the image action mapping results, and then the image action segment corresponding to the next segment anchor is read. Subsequently, based on the termination image boundary information of the image action segment corresponding to the previous segment anchor and the starting image boundary information of the image action segment corresponding to the next segment anchor, it is determined whether there is a continuous image action change on the same time axis between the two image action segments. Content continuity can be determined jointly by image switching markers and action continuation relationships. Image switching markers are used to indicate whether a shot break occurs between adjacent image action segments, and action continuation relationships are used to indicate whether the tail action state of the previous image action segment and the head action state of the next image action segment are continuable states. Action continuation relationships can be read through a pre-configured action continuation table, which consists of continuable image action state pairs and is read along with the target language translation configuration. If the screen transition marker does not indicate a break, and the action continuation table contains records corresponding to the tail state of the previous screen action segment and the head state of the next screen action segment, then the screen continuity status indicates a cross-segment continuation relationship; otherwise, the screen continuity status does not indicate a cross-segment continuation relationship. Once the screen continuity status is determined, it will participate in the generation of the modal dependency identifier along with the original audio continuity status, preventing the restriction of the reading order of parallel carrying channels based solely on the screen continuity status.

[0038] The application obtains the vocal boundaries corresponding to adjacent segment anchor points from the original sound segments and determines the continuity of the original sound based on these boundaries. In this application, the vocal boundary is the boundary information used to describe the start and end states of vocalization in the original sound segments. The vocal boundary includes at least the vocal termination boundary of the original sound segment corresponding to the previous segment anchor point and the vocal start boundary of the original sound segment corresponding to the next segment anchor point. When determining the continuity of the original sound, the vocal termination boundary corresponding to the previous segment anchor point is read first, and then the vocal start boundary corresponding to the next segment anchor point is read. Subsequently, based on the silence interval between the vocal termination boundary and the vocal start boundary, the vocal energy change, and the text pause marker, it is determined whether adjacent original sound segments belong to the same continuous vocalization process. The silence interval is used to represent the silent or low-energy interval between two original sound segments, the vocal energy change is used to represent the energy connection relationship between the preceding and following original sound segments at the boundary, and the text pause marker is used to indicate whether there is a sentence-end pause in the text-carrying segment at the boundary. Text pause markers can be obtained from punctuation, pause marks, or dialogue sentence boundaries in the text-carrying mapping results. If no independently segmentable pause is formed between the sound termination boundary and the sound start boundary, and the text pause marker does not indicate a complete semantic pause, then the original sound continuity state indicates the existence of a cross-segment continuation relationship; if the sound boundary shows that the preceding and following original sound segments have formed independently processable sound intervals, then the original sound continuity state does not indicate the existence of a cross-segment continuation relationship. The original sound continuity state and the image continuity state are used together to generate modal dependency identifiers, so that the modal dependency identifiers reflect both the continuation at the image action segment level and the continuation at the original sound segment level.

[0039] When both the continuous state of the video and the continuous state of the original audio indicate the existence of a cross-segment continuation relationship, a modal dependency identifier pointing to the anchor point of the preceding segment is generated, and this modal dependency identifier is written into the modal track-sharing record corresponding to the current segment anchor point. In this application, the modal dependency identifier is not a conventional marker written to all adjacent segment anchor points, but is only generated when both the continuous state of the video and the continuous state of the original audio indicate the existence of a cross-segment continuation relationship. When generating the modal dependency identifier, the current segment anchor point and the preceding segment anchor point are read first, and then the modal dependency identifier is generated according to the reading direction between the current segment anchor point and the preceding segment anchor point. The modal dependency identifier contains the correspondence between the current segment anchor point and the preceding segment anchor point. The current segment anchor point is used to represent the segment that will enter subsequent parallel processing, and the preceding segment anchor point is used to represent the preceding segment that the current segment anchor point needs to read before entering the parallel carrying channel. After the modal dependency identifier is written into the modal track-sharing record corresponding to the current segment anchor point, it will be read when the decoded inference segment is subsequently captured and used as the dependency carrying field of the decoded inference segment. Therefore, the continuous state of the video and the continuous state of the original sound do not remain independently at the judgment stage, but are transformed into modal dependency identifiers that can be recorded and read by the channel scheduling, so that the cross-segment continuity relationship can directly affect the allocation and reading state of the parallel carrying channel.

[0040] Based on the segment anchor point, the translated reasoning segment is extracted from the modal track-bearing record, including: In this application, when extracting translated inference segments, the film data to be translated is not read directly according to file blocks, but the modal track-based transport record is read first. The modal track-based transport record has established a correspondence between the original sound segments, text transport mapping results, and screen action mapping results according to the segment anchor points. Therefore, reading the segment anchor points in the modal track-based transport record can locate the reloadable time period of the current translated inference segment on the film timeline. At the same time, the modal dependency identifier in the modal track-based transport record is read to determine whether the current segment anchor point needs to retain the cross-segment reading relationship in the subsequent channel scheduling record. If there is no modal dependency identifier in the modal track-based transport record, the translated inference segment corresponding to the current segment anchor point can enter the parallelism judgment according to the cache state; if there is a modal dependency identifier in the modal track-based transport record, the modal dependency identifier will continue to be passed with the translated inference segment, serving as the basis for determining the reading order when generating the subsequent channel scheduling record. By first reading the segment anchor point and modal dependency identifier, the translated reasoning segment can simultaneously possess the re-feedback period and cross-segment reading relationship when it is formed.

[0041] Based on the segment anchor point, the source segment content corresponding to the segment anchor point is extracted from the film data to be translated. In this application, the source segment content is the original content in the film data to be translated defined by the segment anchor point. The source segment content includes at least one or more of the original audio segment, video segment, and original text segment corresponding to the segment anchor point. When extracting the source segment content, the reading range of the source segment content is first determined on a unified timeline based on the segment start time marker and segment end time marker in the segment anchor point. Then, the encapsulation index in the film data to be translated is looked up according to the reading range to obtain the storage location of the original audio data, video data, and original text data in the film data to be translated. Then, the corresponding data block is read according to the storage location to obtain the source segment content. The source segment content is encapsulated in the translation inference segment to provide original audio reference, video reference, and text reference when generating re-translatable content. Since the source segment content is extracted from the film data to be translated based on the segment anchor point, the source segment content carries the time position on a unified timeline consistent with the segment anchor point. It is not necessary to re-establish the time position correspondence when generating the sound duration constraint and video fitting constraint.

[0042] The target language text corresponding to the source segment content is obtained from the target language translation configuration. In this application, the target language text is the translated text corresponding to the source segment content in the target language translation configuration. The target language translation configuration can directly carry the mapping relationship between the segment anchor point and the target language text, or it can carry the mapping relationship between the original text segment and the target language text. When the target language translation configuration carries the mapping relationship between the segment anchor point and the target language text, the target language text is read directly according to the segment anchor point; when the target language translation configuration carries the mapping relationship between the original text segment and the target language text, the original text segment is first read from the source segment content, and then the corresponding target language text is queried in the target language translation configuration according to the original text segment. The target language text is not a separately output subtitle text, but a text input for the subsequent generation of re-translatable content. Therefore, the target language text needs to enter the translation reasoning segment together with the source segment content. After the target language text is read, it will participate in the generation of re-translatable content together with the speech duration constraint and the screen fit constraint, so that the target language text is subject to the constraints of the re-translatable time period and the screen action segment during subsequent processing.

[0043] The application generates a speech duration constraint based on the reloadable time period defined by the segment anchor point. In this application, the reloadable time period is defined by the segment anchor point on a unified time axis, and the speech duration constraint specifies the allowed time range for generating reloadable translated content from the target language text. When generating the speech duration constraint, the segment start time marker and segment end time marker of the segment anchor point are read first, and then the reloadable time period is determined based on these markers. Subsequently, the speech duration constraint is generated based on the reloadable time period. The speech duration constraint may include the speech start boundary, speech end boundary, allowed pause positions, and sentence / paragraph order of the target language text corresponding to the speech start boundary. The speech duration constraint does not simply record a duration, but rather converts the reloadable time period into a constraint field that can be read by subsequent reloadable translated content generation, enabling subsequent parallel transmission channels to read the speech duration constraint when generating the channel translation results, and ensuring that the reloadable translated content falls within the reloadable time period defined by the segment anchor point. In this way, the speech duration constraint transforms the time boundary of the segment anchor point into a time limit that can be directly read within the translated reasoning segment.

[0044] Image fitting constraints are generated based on the image motion mapping results in the modal track-based transport record. In this application, image fitting constraints are used to describe the temporal correspondence between the re-translatable content corresponding to the target language text and the image motion segments. When generating image fitting constraints, the image motion mapping results corresponding to the current segment anchor point are first read from the modal track-based transport record. Then, the image motion segments, the image appearance positions of the image motion segments on a unified time axis, and the image boundary information of the image motion segments are read from the image motion mapping results. Subsequently, image fitting constraints are generated based on the start position, end position, and motion change range of the image motion segments. Image fitting constraints may include the corresponding positions between the vocal units in the target language text and the image motion segments, the start and end boundaries of the image motion segments, and the image motion continuity relationships that need to be retained. Image fitting constraints are read when generating re-translatable content and are used to limit the correspondence between the vocal rhythm of the target language text and the image motion segments. Since the image alignment constraint comes from the image motion mapping result in the modal track carrying record, the image alignment constraint and the segment anchor point maintain the same unified time axis reference, avoiding the repetitive processing introduced by separately aligning the image motion segments after the target text is generated.

[0045] Based on the source segment content, target language text, vocal duration constraints, image fitting constraints, and modal dependency identifier, a translated inference segment is generated, and the modal dependency identifier is used as the dependency carrier field of the translated inference segment. In this application, the translated inference segment is the smallest translation processing unit for subsequent parallel carrier channel reading. When generating the translated inference segment, the source segment content and target language text are first written into the same translated inference segment carrier structure, and then the vocal duration constraints and image fitting constraints are written into the translated inference segment carrier structure, so that the source segment content, target language text, vocal duration constraints, and image fitting constraints maintain a corresponding relationship under the same segment anchor point; then, the modal dependency identifier is read, and the modal dependency identifier is used as the dependency carrier field of the translated inference segment and written into the translated inference segment carrier structure to form the translated inference segment. The dependency carrier field is used to pass the cross-segment reading relationship in the modal track carrier record to the channel scheduling record. If the translated inference fragment carries a dependency carrying field, the subsequent generation of channel scheduling records will determine whether the translated inference fragment can enter the processing slot based on the dependency carrying field. If the translated inference fragment does not carry a dependency carrying field, the subsequent generation of channel scheduling records will mainly rely on the cache state to determine the parallelizability of the translated inference fragment. Thus, the translated inference fragment carries both the content information used to generate the channel translation result and the dependency information used to determine the parallel reading order.

[0046] Parallel bearer channels are allocated to each translated inference segment based on modal dependency identifiers, and channel scheduling records are generated, including: A pre-configured list of parallel bearer channels is obtained. This list is pre-established based on the target language translation configuration and records the channel identifier and cache status of each parallel bearer channel. In this application, the parallel bearer channel list is configuration data used to describe the available parallel bearer channels. The parallel bearer channel list can be established before translation inference begins based on the target language translation configuration. The target language translation configuration can record the target language output format, the types of segments that can be processed simultaneously, the description of available runtime resources, and the result retrieval method. When establishing the parallel bearer channel list based on the target language translation configuration, a channel identifier is assigned to each available parallel bearer channel, and the cache status is recorded for each parallel bearer channel. The channel identifier is used to distinguish different parallel bearer channels in the channel scheduling record, and the cache status is used to indicate whether the parallel bearer channel currently has a cache position to receive translation inference segments. The cache status can be represented as idle, occupied, pending retrieval, or delayed waiting. Once the list of parallel bearer channels is obtained, translation inference fragments will no longer be directly allocated based on the target language translation configuration. Instead, the channel identifiers and cache states in the list of parallel bearer channels will be used to generate channel scheduling records, so that the channel configuration requirements in the target language translation configuration will be transformed into executable scheduling criteria.

[0047] The modal dependency identifier recorded in the dependency carrier field of the translated inference fragment is extracted. In this application, the dependency carrier field is a field written when the translated inference fragment is generated, and it is used to store the modal dependency identifier. When extracting the modal dependency identifier, the dependency carrier field in each translated inference fragment is read, and it is determined whether the dependency carrier field records a modal dependency identifier pointing to the anchor point of the preceding fragment. If the dependency carrier field records a modal dependency identifier, then the modal dependency identifier is used as the reading dependency basis when the current translated inference fragment enters the channel scheduling record; if the dependency carrier field does not record a modal dependency identifier, then the current translated inference fragment does not need to wait for the channel completion identifier of the anchor point of the preceding fragment. The modal dependency identifier is not regenerated here, but is extracted from the translated inference fragment and continues to participate in the allocation of parallel carrier channels. By extracting the modal dependency identifier recorded in the dependency carrier field, the cross-fragment reading relationship generated in the previous processing stage can directly affect the scheduling of parallel carrier channels, instead of re-inferring the cross-fragment relationship in the channel scheduling stage.

[0048] The parallelizability of each translated inference segment is determined based on the cache status and modal dependency identifier. In this application, parallelizability indicates whether the translated inference segment can be immediately written into the pending slot of the parallel bearer channel. To determine parallelizability, the cache status in the parallel bearer channel list is read first, followed by the modal dependency identifier in the translated inference segment. For translated inference segments without modal dependency identifiers, if a parallel bearer channel with an idle cache status exists, the translated inference segment is determined to be in a parallelizable state. For translated inference segments with modal dependency identifiers, it is necessary to further query whether the preceding segment anchor point pointed to by the modal dependency identifier in the channel scheduling record already has a channel completion identifier. If the preceding segment anchor point already has a channel completion identifier and a parallel bearer channel with an idle cache status exists, the translated inference segment is determined to be in a parallelizable state. If the preceding segment anchor point does not have a channel completion identifier, or there is currently no parallel bearer channel with an idle cache status, the translated inference segment is determined not to be in a parallelizable state. Parallelism is determined by both cache state and modal dependency identifier, so that the allocation of parallel bearer channels takes into account both whether the current channel can receive data and whether the decoded inference fragments are constrained by cross-fragment reading relationships.

[0049] In this application, translated inference fragments in a parallelizable state are placed into the pending slot of the corresponding parallel bearer channel, while translated inference fragments not in a parallelizable state are placed into the delay slot of the corresponding parallel bearer channel. The pending slot is a cache location in the parallel bearer channel used to temporarily store readable translated inference fragments, and the delay slot is a cache location in the parallel bearer channel used to temporarily store translated inference fragments not yet allowed to be read. When placing a fragment into a pending slot, a parallel bearer channel with an idle cache status is selected based on the cache status in the parallel bearer channel list. Then, the translated inference fragment in a parallelizable state is written into the pending slot of that parallel bearer channel, and the cache status of that parallel bearer channel is updated to occupied. When placing a fragment into a delay slot, the modal dependency identifier in the translated inference fragment not in a parallelizable state is read, and the translated inference fragment is written into the delay slot of the corresponding parallel bearer channel according to the preceding fragment anchor point pointed to by the modal dependency identifier. The decoded inference segments in the delay slots are not immediately read by the corresponding parallel bearer channels. Instead, processing resumes according to the channel scheduling record after the anchor point of the subsequent preceding segment is marked with a channel completion flag. By separately bearing the pending slots and delay slots, decoded inference segments in parallelizable and non-parallelizable states have different read statuses in the channel scheduling record, enabling the parallel bearer channels to maintain cross-segment read relationships while preserving parallel processing capabilities.

[0050] A channel scheduling record is generated based on the segment anchor points and read states corresponding to the pending slots and delay slots, respectively. In this application, the channel scheduling record is scheduling data jointly generated by the pending slots, delay slots, segment anchor points, read states, channel identifiers, and cache states. When generating the channel scheduling record, the channel identifier of each parallel bearer channel is read first, then the segment anchor point corresponding to the decoded inference segment carried by the pending slot in that parallel bearer channel is read, and the segment anchor point is written into the channel scheduling record at the position corresponding to the channel identifier; subsequently, the segment anchor point corresponding to the decoded inference segment carried by the delay slot is read, and the segment anchor point and modal dependency identifier are written into the channel scheduling record at the position corresponding to the channel identifier. For the decoded inference segment in the pending slot, the channel scheduling record records its read state as allowed to read; for the decoded inference segment in the delay slot, the channel scheduling record records its read state as delayed and waiting. After the channel scheduling record is generated, it will serve as the basis for subsequent parallel bearer channels to read decoded inference segments, and will be used to receive the channel completion identifier after the subsequent channel decoding results are generated. In this way, the channel scheduling record not only records the initial allocation relationship, but also provides a data entry point for subsequent read status updates and delay slot recovery processing.

[0051] like Figure 1 As shown, after generating the modal track-based load record, the translated inference fragment, and the channel scheduling record, subsequent processing no longer re-traverses the movie data to be translated. Instead, the channel scheduling record serves as the direct basis for each parallel load channel to read the translated inference fragment. The channel scheduling record already records the segment anchor points and their reading states corresponding to each parallel load channel. The reading states can include read-allowed state, delayed waiting state, reading in progress state, and completed state. Therefore, the channel scheduling record can pass the refillable time period determined by the segment anchor points, the cross-segment reading relationship determined by the modal dependency identifier, and the cache positions in the pending slot and the delayed slot to the subsequent parallel reading process. In this way, when each parallel load channel reads the translated inference fragment, it does not simply retrieve the fragments according to the order in the file, nor does it arbitrarily read concurrently without dependency judgment. Instead, it performs reading within the segment anchor points and reading states defined by the channel scheduling record, ensuring that the parallel processing of the translated inference fragment is still constrained by the movie timeline and the cross-segment reading relationship.

[0052] Based on the channel scheduling record, each parallel bearer channel reads its corresponding decoded inference segment, including: Based on the segment anchor points and read status in the channel scheduling record, an anchor point gate state is set for each parallel bearer channel. Specifically, in this application, the anchor point gate state is a parallel bearer channel read control state generated based on the channel scheduling record. The anchor point gate state is used to indicate whether a certain parallel bearer channel is currently allowed to read the translated inference segment in the pending slot. When setting the anchor point gate state, the channel identifier in the channel scheduling record is read first, and then the segment anchor point and read status corresponding to the channel identifier are read. Subsequently, the corresponding translated inference segment is determined to be in the pending slot or the delayed slot based on the read status. If the translated inference segment is in the pending slot, the dependency bearer field in the translated inference segment is read, and it is determined whether the preceding segment anchor point corresponding to the modal dependency identifier recorded in the dependency bearer field has been written to the channel completion identifier. If the translated inference segment is in the delayed slot, the segment anchor point corresponding to the translated inference segment, the preceding segment anchor point on which the translated inference segment depends, and the current delayed waiting status of the translated inference segment are written to the anchor point gate state. Therefore, the anchor gatekeeper state is not an independently generated new state, but is jointly determined by the channel identifier, fragment anchor, reading state, and modal dependency identifier in the channel scheduling record, and serves as the direct basis for determining whether reading is allowed when reading the translated inference fragment in the future.

[0053] When the anchor gate status indicates that there are no unmet modal dependency identifiers in the translated inference fragment in the slot to be processed, the corresponding parallel bearer channel is allowed to read the translated inference fragment. Specifically, in this application, an unmet modal dependency identifier refers to a modal dependency identifier pointing to the preceding fragment anchor point in the dependency bearer field, and the channel scheduling record has not yet written a channel completion identifier for the preceding fragment anchor point. When determining whether there are unmet modal dependency identifiers in the translated inference fragment in the slot to be processed, the fragment anchor point corresponding to the translated inference fragment is first read from the anchor gate status, and then the modal dependency identifier is read from the dependency bearer field of the translated inference fragment. If the dependency bearer field does not record a modal dependency identifier, the anchor gate status indicates that there are no unmet modal dependency identifiers in the translated inference fragment. If the dependency bearer field records a modal dependency identifier, the channel completion identifier of the preceding fragment anchor point in the channel scheduling record is queried based on the modal dependency identifier. When the preceding fragment anchor point has been written with a channel completion identifier, the anchor gate status also indicates that there are no unmet modal dependency identifiers in the translated inference fragment. After obtaining the anchor gate status of "allowed reading," the parallel bearer channel reads the translated inference fragment from the corresponding slot according to the channel identifier in the channel scheduling record, and updates the read status of the translated inference fragment from "allowed reading" to "reading in progress." The "reading in progress" status continues to be written to the channel scheduling record, enabling subsequent cache reclamation and delay slot recovery processes to recognize that the translated inference fragment has entered the parallel bearer channel.

[0054] When the preceding segment anchor point on which the translated inference segment in the delay slot depends is marked with a channel completion flag in the channel scheduling record, the anchor point gatekeeper status is updated according to the channel completion flag, and the translated inference segment is transferred from the delay slot to the waiting slot for reading by the corresponding parallel bearer channel. Specifically, in this application, the channel completion flag is written back to the status field of the channel scheduling record after the parallel bearer channel completes the corresponding translated inference segment and generates the channel translation result. The channel completion flag indicates that the corresponding segment anchor point has completed the translation processing required by the current parallel bearer channel. The translated inference segment in the delay slot already carries a modal dependency flag when it enters the delay slot. The modal dependency flag points to the preceding segment anchor point. Therefore, when restoring the translated inference segment in the delay slot, the preceding segment anchor point is first queried in the channel scheduling record according to the modal dependency flag, and then the channel completion flag corresponding to the preceding segment anchor point is read. When a preceding segment anchor is marked with a channel completion flag, it indicates that the preceding segment anchor required by the decoded inference segment in the delay slot has completed its corresponding processing. At this point, the anchor gate status is updated from a delayed waiting state to an allowed reading state based on the channel completion flag. After the anchor gate status is updated, the decoded inference segment is transferred from the delay slot to the pending processing slot, and the reading status of the decoded inference segment in the channel scheduling record is synchronously updated to an allowed reading state. Through this state transition, the delay slot is not merely a cache location for temporarily storing decoded inference segments, but rather works in conjunction with the modal dependency flag, the preceding segment anchor, and the channel completion flag to ensure that decoded inference segments with cross-segment reading relationships resume entering the parallel carrying channel after the preceding segment anchor processing is completed.

[0055] After each parallel carrier channel reads the corresponding translation inference segment, it continues to generate the channel translation result based on the target language text, voice duration constraints, and image fitting constraints in the translation inference segment. Since the translation inference segments originate from the modal track-based carrier record, the segment anchor point, source segment content, target language text, voice duration constraints, image fitting constraints, and modal dependency identifier in the translation inference segment maintain a corresponding relationship within the same translation inference segment carrier structure. Therefore, the parallel carrier channel can directly read these fields when generating the channel translation result, without needing to re-search for the corresponding image action segment or original sound segment from the movie data to be translated. After the channel translation result is generated, it is written to the recycling cache, and the channel completion flag is written back to the channel scheduling record. The channel completion flag not only indicates that the current parallel carrier channel has completed the processing of the corresponding segment anchor point but also triggers the recovery processing of the translation inference segment in the subsequent delay slot.

[0056] The channel translation results include source segment anchors, channel completion markers, re-translated content that can be reloaded, and boundary preservation tags; Each parallel bearer channel generates channel decoding results within the re-feedback period defined by the segment anchor point, including: Each parallel carrier channel reads the corresponding translated reasoning segment within the reloadable time period defined by the segment anchor point. Specifically, in this application, the reloadable time period defined by the segment anchor point originates from the aforementioned modal track-based carrier record and is converted into a sound duration constraint when the translated reasoning segment is formed. When the parallel carrier channel reads the translated reasoning segment, it first reads the segment anchor point in the translated reasoning segment, then reads the reloadable time period defined by the segment anchor point, and subsequently reads the source segment content, target language text, sound duration constraint, and image fitting constraint in the translated reasoning segment within this reloadable time period. The reading here does not only extract the target language text, but also extracts the target language text, the reloadable time period, the image action segment, and the correspondence between the target language text and the image action segment together, so that the subsequent reloadable translated content is already limited by the reloadable time period when it is generated. The reloadable time period here serves as the time boundary for the parallel carrier channel to process the translated reasoning segment, ensuring that the channel translation result does not deviate from the reload range of the segment anchor point on the film timeline.

[0057] Based on the target language text, vocal duration constraints, and image fit constraints in the translated reasoning segment, reloadable translated content is generated. Specifically, in this application, reloadable translated content refers to target language translated content that can be written into the reloadable time period. When generating reloadable translated content, the target language text is first divided into target language vocal units according to the punctuation boundaries and semantic pause positions. A target language vocal unit is a text segment in the target language text that corresponds to a vocal duration range. The target language vocal units will continue to participate in the processing of vocal duration constraints and image fit constraints. Subsequently, the vocal duration constraints are read, and based on the vocal start boundary, vocal end boundary, and allowed pause positions recorded in the vocal duration constraints, a corresponding vocal occupancy range is configured for each target language vocal unit, resulting in a target language vocal arrangement result. The target language vocal arrangement result is used to represent the arrangement order and vocal occupancy range of each target language vocal unit within the reloadable time period. Next, the screen fitting constraints are read, and based on the screen action segments, start and end boundaries of the screen action segments, and screen action continuity relationships recorded in the screen fitting constraints, the target language pronunciation arrangement results are processed to obtain reloadable translation content. The reloadable translation content retains the text content of the target language pronunciation units, while also carrying the pronunciation occupancy range of the target language pronunciation units within the reloadable time period and the corresponding screen action segment position. Thus, the reloadable translation content is not simply translated text data, but rather data content already carried by the reloadable time period and screen action mapping results defined by segment anchor points, which can be directly written into the reloadable time period in subsequent temporal reload calibration.

[0058] Based on the reloadable time period defined by the segment anchor point, a boundary preservation label containing start and end protection points is generated. The start and end protection points indicate the start and end positions of the reloadable time period. Specifically, in this application, the boundary preservation label is used to retain the boundary information of the reloadable time period after the channel translation results enter the recycling cache. When generating the boundary preservation label, the reloadable time period defined by the segment anchor point is first read, and then the start and end positions are determined from the reloadable time period. Subsequently, the start position is written into the start protection point in the start and end protection points, the end position is written into the end protection point in the start and end protection points, and the start and end protection points are written into the boundary preservation label. The start protection point indicates the writable starting point of the reloadable translated content on the movie timeline, and the end protection point indicates the writable ending point of the reloadable translated content on the movie timeline. The boundary preservation label can also carry the correspondence between the segment anchor point and the reloadable time period, so that the channel translation results belonging to the same segment anchor point in the recycling cache can read the same boundary criteria when entering the recycling readiness judgment. Since the boundary preservation label is generated by the re-feedback period defined by the fragment anchor point, the boundary preservation label can limit the channel translation results to the time boundary of the original fragment anchor point in subsequent consistency verification and timing re-feedback calibration.

[0059] The fragment anchor point is written as a same-source fragment anchor point into the channel translation result, and a channel completion identifier is written into the channel translation result. Specifically, in this application, the same-source fragment anchor point is used to indicate which fragment anchor point the channel translation result originates from. After generating re-fedable translated content and boundary preservation tags, the parallel bearer channel reads the fragment anchor point in the translated inference fragment and writes it as a same-source fragment anchor point into the channel translation result. Since the same-source fragment anchor point is the same as the fragment anchor point in the translated inference fragment, subsequent cache retrieval can directly assign the channel translation result to the corresponding index position according to the same-source fragment anchor point. The channel completion identifier is used to indicate that the parallel bearer channel has completed reading the translated inference fragment and generating the channel translation result. When writing the channel completion identifier, it is first confirmed that the channel translation result already contains re-fedable translated content, boundary preservation tags, and same-source fragment anchor points, and then the channel completion identifier is written into the channel translation result. The channel completion identifier is then written back to the channel scheduling record to trigger the recovery processing of other translated inference fragments that depend on this fragment anchor point. By combining the same-source fragment anchor and the channel completion identifier, the channel translation result can be both recycled and cached according to the fragment anchor, and can also provide a basis for subsequent reading status updates for the channel scheduling record.

[0060] The channel translation result carrying the same-source fragment anchor point is written to the recycling cache, and the channel completion flag is written back to the channel scheduling record. Specifically, in this application, the recycling cache is indexed by the fragment anchor point. When the channel translation result is written to the recycling cache, the same-source fragment anchor point in the channel translation result is read first, and then the index position in the recycling cache is determined according to the same-source fragment anchor point. If a cache position corresponding to the same-source fragment anchor point already exists in the recycling cache, the channel translation result is written to that cache position; if a cache position corresponding to the same-source fragment anchor point does not yet exist in the recycling cache, the corresponding cache position is established according to the same-source fragment anchor point, and the channel translation result is written to that cache position. After the channel translation result is written to the recycling cache, the channel completion flag in the channel translation result is read, and the channel completion flag is written back to the position in the channel scheduling record associated with the same-source fragment anchor point and the corresponding parallel bearer channel. After the channel completion flag is written back to the channel scheduling record, the read status in the channel scheduling record is updated to the completed status. This completed status will be read when the channel recycling conditions are judged and the translated inference fragments in the recovery delay slot are restored. Through this two-way connection between the cache and the channel scheduling record, the channel translation results are not directly written back to the movie timeline, but the results are collected and the status is written back before entering the timing backflow calibration.

[0061] The decoded channel results belonging to the same segment anchor point in the reclamation cache were detected to meet the reclamation readiness conditions, including: Taking the channel translation results corresponding to the same-source segment anchors in the recycling cache as the object, the registered parallel bearer channels are queried in the channel scheduling record based on the same-source segment anchors. Specifically, in this application, the registered parallel bearer channels are those parallel bearer channels in the channel scheduling record that have a corresponding relationship with the same-source segment anchors. When determining the recycling readiness condition, the channel translation results corresponding to the same-source segment anchors are first read from the recycling cache, and then the channel scheduling record is queried based on the same-source segment anchors. The channel scheduling record already records the segment anchors corresponding to each parallel bearer channel and their reading status. Therefore, all parallel bearer channels registered under the same segment anchor can be queried through the same-source segment anchor. The queried registered parallel bearer channels are used to determine whether there are any unfinished parallel processes under the same segment anchor. If a parallel bearer channel has been registered under the same-source segment anchor, but the reading status corresponding to the parallel bearer channel has not been updated to the completed status, the channel translation results corresponding to the same-source segment anchor cannot enter the timing backfeed calibration. Therefore, the registered parallel bearer channels will associate the channel translation results in the recycling cache with the read status in the channel scheduling record, so that the recycling readiness condition can reflect whether the channel translation results under the same segment anchor point have been collected.

[0062] When all registered parallel bearer channels have channel completion identifiers, the channel recycling condition is confirmed to be met. Specifically, in this application, the channel recycling condition indicates that all registered parallel bearer channels under the same segment anchor point have completed the processing of the corresponding translated inference segment. When confirming the channel recycling condition, the channel completion identifiers of each registered parallel bearer channel in the channel scheduling record are read one by one. If every registered parallel bearer channel has a channel completion identifier, the channel recycling condition is confirmed to be met. If at least one registered parallel bearer channel does not have a channel completion identifier, the channel recycling condition is confirmed not to be met, and the channel translation results of the corresponding same-source segment anchor point are kept in the recycling cache. The channel recycling condition can prevent some channel translation results under the same segment anchor point from entering the timing backfill calibration first, thus ensuring that the channel translation results in the recycling cache have complete channel completion identifiers before entering the movie timeline. After the channel recycling condition is confirmed, it will participate in the determination of the recycling readiness condition together with the consistency check of the boundary maintenance tag and the content backfill condition.

[0063] When the end protection point of the preceding segment anchor and the start protection point of the following segment anchor are contiguous on the timeline, as indicated by the boundary preservation labels of adjacent segment anchors, the boundary preservation label is confirmed to have passed the consistency check. Specifically, in this application, the consistency check is used to determine whether the reloadable time period between adjacent segment anchors maintains a contiguous relationship on the movie timeline. During the consistency check, the boundary preservation label in the translation result of the channel corresponding to the preceding segment anchor is read first, and then the boundary preservation label in the translation result of the channel corresponding to the following segment anchor is read. Subsequently, the end protection point is read from the boundary preservation label of the preceding segment anchor, and the start protection point is read from the boundary preservation label of the following segment anchor. If the end protection point of the preceding segment anchor and the start protection point of the following segment anchor are contiguous on the movie timeline, the boundary preservation label is confirmed to have passed the consistency check. If there is an overlap or break between the end protection point of the preceding segment anchor and the start protection point of the following segment anchor that is not interpreted by the corresponding segment anchor, the boundary preservation label is confirmed to have failed the consistency check, and the relevant channel translation result continues to be kept in the reclaimed cache. After the boundary preservation label passes the consistency check, it means that the channel translation results of adjacent segment anchor points have a continuous start and end protection point when entering the time-series reflow calibration, and the subsequent reflowable translation content can enter the reflowable period according to the boundary preservation label.

[0064] When the reloadable translated content meets the speech duration constraint and screen fit constraint, the content reload condition is confirmed to be met. Specifically, in this application, the content reload condition is used to determine whether the reloadable translated content is suitable for inclusion in the reloadable time period defined by the segment anchor point. When confirming the content reload condition, the reloadable translated content is first read from the channel translation results, and then the speech duration constraint and screen fit constraint used when generating the reloadable translated content are read. For the speech duration constraint, the speech occupancy range of each target language speech unit in the reloadable translated content is read, and it is determined whether the speech occupancy range of each target language speech unit falls within the reloadable time period. For the screen fit constraint, the screen action segment position corresponding to each target language speech unit in the reloadable translated content is read, and it is determined whether the speech occupancy range of the target language speech unit corresponds to the screen action segment position. If the reloadable translated content meets both the audio duration constraint and the image fit constraint, then the content reload condition is confirmed to be met; if the reloadable translated content does not meet either the audio duration constraint or the image fit constraint, then the content reload condition is confirmed to be unmet. The content reload condition ensures that the time constraint and image constraint are read and judged before the reloadable translated content enters the movie timeline, thereby reducing the number of boundary adjustments performed after the reloadable translated content is written.

[0065] The recycling readiness condition is determined to be satisfied when all three conditions—channel recycling, consistency verification, and content refilling—are met. Specifically, in this application, the recycling readiness condition is composed of the channel recycling condition, consistency verification, and content refilling condition. When determining the recycling readiness condition, the confirmation result of the channel recycling condition is read first, followed by the confirmation result of the consistency verification, and then the confirmation result of the content refilling condition. The recycling readiness condition is only determined to be satisfied when all three conditions are met. The channel recycling condition constrains the completion status of parallel carrier channels under the same segment anchor point; the consistency verification constrains the boundary connection status between adjacent segment anchor points; and the content refilling condition constrains the correspondence between refillable translated content and the audio duration constraint and image fit constraint. The confirmation results of the channel recycling condition, the consistency verification, and the content refilling condition correspond to the completion status of the parallel carrier channel, the boundary connection status between adjacent segment anchor points, and the content refilling status of the refillable translated content, respectively. Only after all the aforementioned confirmation results are met does the channel translation result in the recycling cache enter the timing refilling calibration.

[0066] If any condition is not met, the corresponding channel translation result is kept in the recycling cache, and the channel translation result is prohibited from entering the time-series backflow calibration. Specifically, in this application, when any condition is not met, the corresponding channel translation result will not be deleted, nor will it be directly written back to the movie timeline, but will continue to be kept in the recycling cache. The channel translation result kept in the recycling cache is still indexed by the same source segment anchor point, and continues to wait for the channel to complete the identifier write-back, the boundary preservation label update, or the re-entry of the re-translated content into the content backflow condition judgment. Prohibiting the channel translation result from entering the time-series backflow calibration means that before the recycling readiness condition is met, no write position is allocated to the channel translation result on the movie timeline, nor is the calibrated anchor point state generated based on the channel translation result. Through this processing, the recycling cache is not only used to store the channel translation result, but also to prevent the channel translation result that has not met the recycling readiness condition from entering the time-series backflow calibration in advance, so that the results generated by multiple channels in parallel can complete the necessary state judgment before being written to the movie timeline.

[0067] Perform timing backfeed calibration on the channel translation results, including: Taking the channel translation results that meet the retrieval readiness conditions as the object, the corresponding reloadable time period on the timeline is determined based on the source segment anchor points. Specifically, in this application, the channel translation results that meet the retrieval readiness conditions have passed the judgment of channel retrieval conditions, consistency verification, and content reload conditions, and therefore can enter the time-series reload calibration. When performing time-series reload calibration, the source segment anchor points are first read from the channel translation results, and then the segment anchor points in the modal track carrying record are retrieved based on the source segment anchor points. The corresponding reloadable time period on the movie timeline is then read based on the segment anchor points. Since the source segment anchor points originate from the segment anchor points in the translated inference segments, the source segment anchor points are consistent with the segment anchor points in the modal track carrying record. Determining the reloadable time period based on the source segment anchor points allows the time-series reload calibration to use the same time boundary when cutting the translated inference segments, avoiding deviation of the channel translation results from the original segment anchor points when writing them onto the movie timeline.

[0068] The start and end protection points of the re-feedback period are determined based on boundary preservation tags. Specifically, in this application, the boundary preservation tags are written during the channel translation result generation stage, and the start and end protection points in the boundary preservation tags originate from the re-feedback period defined by the segment anchor points. After determining the re-feedback period, the boundary preservation tags in the channel translation results are read, and the start and end protection points are read from the boundary preservation tags. The start protection point is used to determine the starting position of the re-feedback translated content written into the re-feedback period, and the end protection point is used to determine the ending position of the re-feedback translated content written into the re-feedback period. The start and end protection points, together with the re-feedback period, participate in defining the writing range, where the re-feedback period provides the overall time range, and the start and end protection points provide the writing boundary positions. In this way, the timing re-feedback calibration is not written into the movie timeline solely based on the generation order of the channel translation results, but rather the writing range is determined within the re-feedback period based on the boundary preservation tags.

[0069] Within the range defined by the start and end protection points, rewritable translated content is written into the rewritable time period. Specifically, in this application, the rewritable translated content already meets the sound duration constraints and screen fit constraints before writing; therefore, the writing process mainly determines the writing range based on the start and end protection points. When writing rewritable translated content, the start and end protection points are read first, and then the writing range defined by the start and end protection points is determined on the movie timeline. Subsequently, the sound occupation range of each target language sound unit in the rewritable translated content is read, and each target language sound unit is written into the writing range according to the target language sound arrangement result. During the writing process, the starting position of the rewritable translated content is no earlier than the start protection point, and the ending position of the rewritable translated content does not exceed the end protection point. If the rewritable translated content carries a target language sound unit corresponding to the position of the screen action segment, then according to the correspondence between the target language sound unit and the position of the screen action segment, the target language sound unit is placed into the corresponding sound occupation range in the rewritable time period. This process allows re-translated content to be written into the re-translatable time slot in a manner that is jointly defined by segment anchors, boundary preservation tags, and screen fit constraints, rather than being inserted without boundaries on the movie timeline.

[0070] Based on the reloadable time period that has been written, a calibrated anchor point state is generated, and this calibrated anchor point state is written back to the channel scheduling record. This allows the translation inference segment in the delay slot to resume processing based on the calibrated anchor point state, thus obtaining the movie translation inference result. Specifically, in this application, the reloadable time period that has been written refers to the movie timeline interval after the reloadable translation content has been written according to the start and end protection points. When generating the calibrated anchor point state, the reloadable time period that has been written is read first, followed by the write position, start protection point, end protection point, and source segment anchor point of the reloadable translation content. This information is then written into the calibrated anchor point state. The calibrated anchor point state indicates that the channel translation result corresponding to the source segment anchor point has completed timing reload calibration, and the reloadable time period corresponding to the source segment anchor point is available for subsequent dependent segments to read. After the calibrated anchor point state is written back to the channel scheduling record, the read status corresponding to the source segment anchor point in the channel scheduling record is updated, and the translation inference segment in the delay slot can query the calibrated anchor point state based on the modal dependency identifier. When the calibrated anchor point status corresponding to the preceding segment anchor point on which the translated inference segment in the delay slot depends has been written into the channel scheduling record, the translated inference segment can be transferred from the delay slot to the pending slot and continue to be read into the parallel bearer channel. The movie translation inference result consists of the re-feedable translated content that has completed time-series re-feedback calibration, the corresponding re-feedback time period, and the calibrated anchor point status. The movie translation inference result can reflect the re-feedback position of the translated content on the movie timeline and the continuity status between adjacent segment anchor points.

[0071] In an easily understood embodiment, the film data to be translated contains two adjacent dialogue segments. The first pair of dialogue segments continues to have audio action at the end of its on-screen action segment, while the second pair of dialogue segments' original audio action segment immediately follows the audio boundary of the first pair. The on-screen continuity state, obtained from the continuity of the on-screen action segment, indicates a cross-segment continuation relationship. Similarly, the original audio continuity state, obtained from the audio boundary, also indicates a cross-segment continuation relationship. Therefore, a modal dependency identifier pointing to the first pair of dialogue segments is written into the modal track-based bearer record corresponding to the second pair of dialogue segments. After the second pair of dialogue segments generates a translation inference segment, this inference segment, carrying the modal dependency identifier, first enters the delay slot. The parallel bearer channel corresponding to the first pair of dialogue segments generates the channel translation result and writes the channel completion identifier back to the channel scheduling record. The anchor gatekeeper state of the second pair of dialogue segments is then updated, and the translation inference segment of the second pair of dialogue segments is transferred from the delay slot to the pending processing slot. Subsequently, the parallel bearer channel reads the translation inference segment of the second pair of dialogue segments and generates the channel translation result within the re-feedback period of the second pair of dialogue segments. After the channel translation results of the two dialogue segments are written to the recycling cache, consistency verification is performed using the start and end protection points in the boundary preservation tags. Once the channel recycling conditions are met, the boundary preservation tags pass the consistency verification, and the re-translated content meets the content re-recycling conditions, the channel translation results enter the time-series re-recycling calibration. This embodiment illustrates that this application does not simply distribute movie segments to multiple parallel carrying channels, but rather transmits the status information corresponding to the same segment anchor point between delay slots, anchor point gate status, channel completion flags, recycling cache, and time-series re-recycling calibration, ensuring that cross-segment reading relationships and re-recyclable time periods continuously exert constraints during parallel processing.

[0072] Through the aforementioned processing, the film data to be processed and the translation constraint configuration can also be understood in a higher-level way as a summary of the film data to be translated and the target language translation configuration. Multimodal temporal relationships can correspond to the presentation timing of various heterogeneous modal content; temporal carrying information can correspond to modal track carrying records; processing boundaries can correspond to the reloadable time periods defined by segment anchor points; reading dependencies can correspond to modal dependency identifiers; translation inference carrying objects can correspond to translation inference segments; parallel execution guidelines can correspond to channel scheduling records; multiple parallel processing paths can correspond to each parallel carrying channel; path translation results can correspond to channel translation results; the result recycling area can correspond to the recycling cache; recycling judgment conditions can correspond to recycling ready conditions; and temporal calibration processing can correspond to temporal reload calibration. After the film data to be processed and the translation constraint configuration enter the above processing, the processing boundaries and reading dependencies are determined through temporal carrying information, and then the translation inference carrying objects are made to enter multiple parallel processing paths through parallel execution guidelines; after multiple parallel processing paths generate path translation results, the path translation results enter the result recycling area and enter temporal calibration processing according to the recycling judgment conditions.

[0073] In a specific application scenario, the film data to be translated can be a film file containing video data, original audio data, raw text data, and a wrapper index (e.g., the video file has a frame rate of 24 frames per second, the original audio data is sampled at 48,000 times per second, and the raw text data is the subtitle track wrapped with the film file). The target language translation configuration can be a translation task configuration file corresponding to the film data to be translated. The target language translation configuration records the reading position of the target language text, the correspondence between the target language text and the original text data, the number of parallel carrying channels, and the writing method of the cache (e.g., the number of parallel carrying channels can be configured to 4, and the writing method of the cache can be configured to establish the cache position according to the segment anchor point). When reading the film data to be translated, the timestamp information corresponding to the video data, original audio data, and raw text data is first read from the wrapper index, and then these timestamp information are converted to the same film timeline. The converted movie timeline can be expressed in milliseconds (for example, starting from the beginning of the movie playback, a certain screen frame corresponds to the 10200th millisecond, a certain audio frame corresponds to the 10208th millisecond, and a certain original text data corresponds to the 10180th to 11400th milliseconds). The converted movie timeline continues to be used to mark the original sound segments, text segments, and screen action segments.

[0074] In this application scenario, assume that an original sound segment in the film data to be translated is located between 10000 and 12600 milliseconds on the film's timeline. This original sound segment is determined by the audio frame energy changes and silence intervals in the original audio data. The corresponding text segment in the original text data is located between 10040 and 12580 milliseconds on the film's timeline, and the corresponding motion segment in the video data is located between 9980 and 12620 milliseconds on the film's timeline. Based on the position of the text segment on the film's timeline, the text segment is mapped to the corresponding segment anchor point of the original sound segment, resulting in a text segment mapping result. Similarly, based on the position of the motion segment on the film's timeline, the motion segment is mapped to the corresponding segment anchor point of the same original sound segment, resulting in a motion mapping result. The text-carrying mapping result carries the content storage location of the text-carrying segment, the position of the text-carrying segment on the movie timeline, and the corresponding segment anchor point. The image-motion mapping result carries the position of the image-motion segment on the movie timeline, the image boundary information of the image-motion segment, and the corresponding segment anchor point. Subsequently, a modal track-based carrying record is generated based on the text-carrying mapping result and the image-motion mapping result, so that the segment anchor point is simultaneously associated with the original sound segment, the text-carrying segment, and the image-motion segment. At this point, the modal track-based carrying record does not simply record a time period, but rather records the rewritable time period of the segment anchor point on the movie timeline, the text-carrying segment corresponding to the segment anchor point, and the image-motion segment corresponding to the segment anchor point.

[0075] If the original sound segment of the preceding segment anchor point is located between 7600 and 10020 milliseconds on the movie timeline, and the original sound segment of the current segment anchor point starts at 10000 milliseconds on the movie timeline, and the end of the screen action segment corresponding to the preceding segment anchor point and the beginning of the screen action segment corresponding to the current segment anchor point are continuous in terms of screen action, then the screen continuity state is determined based on the screen action segments corresponding to the preceding and current segment anchor points. If there is no independently segmentable silence interval between the sound termination boundary of the preceding segment anchor point and the sound start boundary of the current segment anchor point, and the text pause marker in the text carrier mapping result does not indicate a complete pause, then the original sound continuity state is determined based on the sound boundary corresponding to the preceding and current segment anchor points. When both the screen continuity state and the original sound continuity state indicate the existence of a cross-segment continuation relationship, a modal dependency identifier pointing to the preceding segment anchor point is generated, and this modal dependency identifier is written into the modal track carrier record corresponding to the current segment anchor point. In this way, before the translated inference fragment corresponding to the current fragment anchor point enters the parallel carrying channel, it needs to confirm whether the previous fragment anchor point has been written into the channel completion flag based on the modal dependency flag. After the channel translation result corresponding to the previous fragment anchor point completes the timing backfeed calibration, the translated inference fragment corresponding to the current fragment anchor point can also read the calibrated anchor point status corresponding to the previous fragment anchor point based on the modal dependency flag.

[0076] When generating a translated inference fragment, the modal track carrying record corresponding to the current fragment anchor point is first read. Then, the source fragment content is extracted from the film data to be translated based on the current fragment anchor point. The source fragment content may include the original audio fragment, video fragment, and original text fragment corresponding to the current fragment anchor point. Subsequently, the target language text corresponding to the source fragment content is read according to the target language translation configuration (for example, the target language text can be read from the mapping relationship between the fragment anchor point and the target language text in the target language translation configuration). A voice duration constraint is generated based on the refeedable time period defined by the current fragment anchor point. The voice duration constraint can record the allowed start position, allowed end position, and allowed pause position of the target language text. Then, a screen fitting constraint is generated based on the video action mapping result. The screen fitting constraint can record the start and end boundaries of the video action segment and the corresponding position between each target language voice unit in the target language text and the video action segment. After the source fragment content, target language text, voice duration constraint, screen fitting constraint, and modal dependency identifier are written into the same translated inference fragment carrying structure, a translated inference fragment is formed. The translated reasoning fragment includes source fragment content, target language text, speech duration constraints, and screen fitting constraints used to generate re-translatable content, as well as a dependency carrier field used to determine the reading order. The dependency carrier field is used to record modal dependency identifiers.

[0077] During the parallel bearer channel allocation process, a pre-configured list of parallel bearer channels can record the channel identifiers and cache statuses of four parallel bearer channels. If the first and second parallel bearer channels are idle, the third parallel bearer channel is occupied, and the fourth parallel bearer channel is in a delayed waiting state, then when processing a translated inference fragment without a modal dependency identifier, the translated inference fragment can be written to the pending slot of the first or second parallel bearer channel. For translated inference fragments carrying modal dependency identifiers, it is necessary to first query the channel completion identifier of the preceding fragment anchor point in the channel scheduling record based on the modal dependency identifier. If the preceding fragment anchor point has not yet been written to the channel completion identifier, the translated inference fragment is written to the delayed slot of the corresponding parallel bearer channel; if the preceding fragment anchor point has already been written to the channel completion identifier, and there is an idle parallel bearer channel, the translated inference fragment is written to the pending slot of the corresponding parallel bearer channel. The segment anchor point, channel identifier, and reading status corresponding to the pending slot and the delay slot are written into the channel scheduling record, which then becomes the basis for subsequent reading of the translated inference segment.

[0078] When reading translated inference segments in the parallel bearer channel, an anchor gate state is first set for the parallel bearer channel according to the channel scheduling record. If a translated inference segment is located in the pending slot and there are no unmet modal dependency flags for that segment, the anchor gate state is set to the allowed read state, and the parallel bearer channel reads that translated inference segment. If a translated inference segment is located in the delayed slot and the modal dependency flag carried by that segment points to the anchor point of the preceding segment, the anchor gate state is set to the delayed waiting state. After the channel translation result of the preceding segment anchor point is generated, the channel completion flag is written back to the channel scheduling record, and the anchor gate state is updated from the delayed waiting state to the allowed read state based on the channel completion flag. After the anchor gate state is updated, the translated inference segment is moved from the delayed slot to the pending slot, and the read state of the translated inference segment in the channel scheduling record is synchronously updated to the allowed read state. Through this cache position conversion, consecutive segments on the same movie timeline will not prematurely read segments with incomplete dependencies due to parallel processing.

[0079] When generating the channel translation result, the parallel channel reads the target language text, vocal duration constraints, and screen fitting constraints from the translated inference segment. The target language text can be divided into several target language vocal units (e.g., the target language text can be divided into 3 target language vocal units according to sentence pauses). The vocal duration constraint can configure a corresponding vocal occupancy range for each target language vocal unit, resulting in the target language vocal arrangement result. Subsequently, screen fitting processing is performed on the target language vocal arrangement result according to the screen fitting constraints, so that the vocal occupancy range of the target language vocal unit corresponds to the position of the screen action segment, thereby generating re-translatable content. At the same time, boundary preservation labels are generated based on the re-translatable time period defined by the segment anchor point. The start protection point in the boundary preservation label can correspond to the start position of the re-translatable time period, and the end protection point in the boundary preservation label can correspond to the end position of the re-translatable time period. Then, the segment anchor point is written as the same source segment anchor point into the channel translation result, and the channel completion identifier is written into the channel translation result. The channel translation result consists of the same source segment anchor, channel completion identifier, re-translatable content, and boundary preservation tag. After the channel translation result is written to the recycling cache, the channel completion identifier is synchronously written back to the channel scheduling record.

[0080] When determining the readiness conditions for recycling, the channel scheduling record is first queried based on the same-source segment anchor point in the recycling cache to read the registered parallel bearer channels under that same-source segment anchor point. If two parallel bearer channels are registered under that same-source segment anchor point, both parallel bearer channels must have channel completion identifiers to confirm that the channel recycling conditions are met. Then, the boundary preservation tags corresponding to adjacent segment anchor points are read, and the continuity relationship between the termination protection point of the preceding segment anchor point and the start protection point of the subsequent segment anchor point on the movie timeline is compared. If the termination protection point of the preceding segment anchor point corresponds to the 12600th millisecond, and the start protection point of the subsequent segment anchor point also corresponds to the 12600th millisecond, or the interval between the two can be explained by the pause position in the corresponding segment anchor point, then the boundary preservation tag is confirmed to have passed the consistency check. Then, the reloadable translated content, the voice duration constraint, and the image fit constraint are read to determine whether the target language voice unit in the reloadable translated content falls within the reloadable time period defined by the voice duration constraint, and whether the target language voice unit maintains the correspondence with the image action segment position corresponding to the image fit constraint. When the channel reclamation conditions, consistency verification, and content reflow conditions are all met, the channel translation results in the reclamation cache will enter the time-series reflow calibration; if any condition is not met, the channel translation results will remain in the reclamation cache and will not enter the movie timeline.

[0081] During timing backfeed calibration, the corresponding backfeedable time slot on the movie timeline is first determined based on the source segment anchor points in the channel translation results. Then, the start and end protection points of the backfeedable time slot are determined based on the boundary preservation labels. Subsequently, within the write range defined by the start and end protection points, the backfeedable translated content is written to the backfeedable time slot. After writing, the calibrated anchor point state is generated based on the completed backfeedable time slot, the write position of the backfeedable translated content, the start and end protection points, and the source segment anchor points. After the calibrated anchor point state is written back to the channel scheduling record, the read status of the corresponding source segment anchor point in the channel scheduling record is updated. If the translated inference segment in the subsequent delay slot carries a modal dependency identifier pointing to the source segment anchor point, the translated inference segment in the subsequent delay slot can be restored based on the calibrated anchor point state. Through the above visualization process, the picture data, original audio data and original text data in the movie data to be translated can form a modal track carrying record on the movie timeline. The translated inference segment can enter the processing slot or delay slot according to the modal dependency identifier. The channel translation result can be written to the recycling cache first and then enter the time sequence backflow calibration, and finally the movie translation inference result is obtained.

[0082] like Figure 3 As shown, this embodiment of the present application provides a movie dubbing inference device that supports multimodal parallel processing, comprising: The modal track-based delivery module is used to acquire the film data to be translated and the target language translation configuration, and generate modal track-based delivery records according to the presentation sequence of multiple heterogeneous modal content in the film data to be translated. The modal track-based delivery record includes segment anchors and modal dependency identifiers. The segment anchors are used to define the re-feedable time period of the translated content on the film timeline, and the modal dependency identifiers are used to represent the cross-segment reading relationship between different segment anchors. The channel scheduling record generation module is used to extract the translated inference segment from the modal track-based bearer record according to the segment anchor point, and allocate parallel bearer channels for each translated inference segment based on the modal dependency identifier, thereby generating a channel scheduling record; the channel scheduling record records the segment anchor point and its reading status corresponding to each parallel bearer channel; The parallel bearer execution module is used to drive each of the parallel bearer channels to read the corresponding translation inference fragments according to the channel scheduling record, so that each parallel bearer channel generates the channel translation result within the recycle period defined by the fragment anchor point, and writes the channel translation result into the recycling cache indexed by the fragment anchor point. The timing backfeed calibration module is used to perform timing backfeed calibration on the channel translation results when the channel translation results belonging to the same segment anchor point in the recycling cache meet the recycling ready conditions, so as to obtain the movie translation inference results.

[0083] like Figure 4 As shown, an electronic device according to an embodiment of this application includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it causes the processor to execute the movie translation inference method supporting multimodal parallel processing as described in any one of the claims of this application.

[0084] like Figure 5 As shown, a computer-readable storage medium is provided in an embodiment of this application. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the movie translation reasoning method supporting multimodal parallel processing as described in any one of the claims of this application.

[0085] like Figure 6 As shown, this is an embodiment of a movie translation inference system that supports multimodal parallel processing, which includes a movie data access node, a channel scheduling node, a parallel carrying node, and a recycling and recharge node. The film data access node is used to acquire the film data to be translated and the target language translation configuration. Based on the presentation sequence of multiple heterogeneous modal content in the film data to be translated, it generates a modal track-based carrying record and sends the modal track-based carrying record to the channel scheduling node. The modal track-based carrying record includes segment anchors and modal dependency identifiers. The segment anchors are used to define the re-feedable time period of the translated content on the film timeline, and the modal dependency identifiers are used to represent the cross-segment reading relationship between different segment anchors. The channel scheduling node is used to extract the translated inference segment from the modal track-based bearer record according to the segment anchor point, allocate parallel bearer channels for each translated inference segment based on the modal dependency identifier, generate a channel scheduling record, and send the channel scheduling record to the parallel bearer node; the channel scheduling record records the segment anchor point and its reading status corresponding to each parallel bearer channel; The parallel bearer node is used to drive each of the parallel bearer channels to read the corresponding translation inference fragments according to the channel scheduling record, so that each parallel bearer channel generates the channel translation result within the recycle period defined by the fragment anchor point, and writes the channel translation result into the recycling cache indexed by the fragment anchor point. The recycling and recharge node is used to perform time-series recharge calibration on the channel translation results when the channel translation results belonging to the same segment anchor point in the recycling cache meet the recycling readiness conditions, so as to obtain the movie translation inference results.

[0086] like Figure 7 As shown, this embodiment of the present application provides a movie translation inference method supporting multimodal parallel processing, which includes: Obtain the movie data to be processed and the translation constraint configuration, and generate time-carrying information based on the multimodal temporal correlation in the movie data to be processed. The time-carrying information is used to limit the processing boundary of the translated reasoning content in the movie time sequence and the reading dependency relationship between different processing boundaries. A translation inference bearer object is constructed based on the timing bearer information, and a parallel execution guide is configured for the translation inference bearer object according to the reading dependency relationship. The parallel execution guide is used to indicate the bearer position and reading status of each translation inference bearer object. According to the parallel execution guidelines, multiple parallel processing paths are driven to read the corresponding translation inference bearer objects, so that the multiple parallel processing paths generate path translation results under the processing boundary constraints, and write the path translation results into the result recycling area; When the path translation results in the result recovery area meet the recovery judgment conditions used to characterize the recirculation of multiple path results within the same processing boundary, the path translation results are subjected to time-series calibration processing to obtain the movie translation inference results.

[0087] The above Figures 3-7 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A movie dubbing inference method supporting multimodal parallel processing, characterized in that, include: Acquire the film data to be translated and the target language translation configuration, and generate a modal track carrying record based on the presentation sequence of multiple heterogeneous modal content in the film data to be translated; Based on the segment anchor point, the translated inference segment is extracted from the modal track-based bearer record, and parallel bearer channels are allocated to each translated inference segment based on the modal dependency identifier, generating a channel scheduling record; Based on the channel scheduling record, each parallel bearer channel is driven to read the corresponding translation inference fragment, so that each parallel bearer channel generates the channel translation result within the recycle period defined by the fragment anchor point, and writes the channel translation result into the recycling cache indexed by the fragment anchor point. When the channel translation results belonging to the same segment anchor point in the recycling cache are detected to meet the recycling readiness conditions, the channel translation results are subjected to time-series backflow calibration to obtain the movie translation inference results.

2. The method according to claim 1, characterized in that, Based on the presentation sequence of various heterogeneous modal content in the film data to be translated, a modal track carrying record is generated, including: Parse the encapsulation index in the movie data to be translated, and establish a unified timeline based on the timestamp information in the encapsulation index; Based on the presentation sequence of heterogeneous modal content, the original sound segments are marked on a unified timeline; The text carrier segment is mapped to the segment anchor point corresponding to the original sound segment to obtain the text carrier mapping result, which carries the segment anchor point corresponding to the text carrier segment. Map the on-screen action segment to the corresponding segment anchor point of the original sound segment to obtain the on-screen action mapping result, which carries the segment anchor point corresponding to the on-screen action segment. Modal track-based carrying records are generated based on the text carrying mapping results and the screen motion mapping results. When there is a cross-segment continuation relationship, a modal dependency identifier is written to the modal track-based carrying record.

3. The method according to claim 2, characterized in that, Modal dependency identifiers are generated based on the continuity of the visuals and the continuity of the original audio at adjacent segment anchor points, including: Obtain the video action segments corresponding to the anchor points of adjacent segments from the video action mapping results, and determine the video continuity state based on the content continuity of the video action segments; Obtain the sound boundaries corresponding to the anchor points of adjacent segments from the original sound segments, and determine the continuity state of the original sound based on the sound boundaries; When both the continuous state of the video and the continuous state of the original sound indicate the existence of a cross-segment continuation relationship, a modal dependency identifier pointing to the anchor point of the previous segment is generated, and the modal dependency identifier is written into the modal track carrying record corresponding to the current segment anchor point.

4. The method according to claim 2, characterized in that, Based on the segment anchor point, the translated reasoning segment is extracted from the modal track-bearing record, including: Read the segment anchor points and modal dependency identifiers from the modal track bearing record; Based on the segment anchor point, extract the source segment content corresponding to the segment anchor point from the movie data to be translated; Obtain the target language text corresponding to the source fragment content from the target language translation configuration; The sound duration constraint is generated based on the rechargeable time period defined by the segment anchor point; Image fitting constraints are generated based on the image motion mapping results in the modal track splitting load record. Based on the source fragment content, target language text, speech duration constraints, screen fit constraints, and modal dependency identifier, a translated reasoning fragment is generated, and the modal dependency identifier is used as the dependency carrier field of the translated reasoning fragment.

5. The method according to claim 4, characterized in that, Parallel bearer channels are allocated to each translated inference segment based on modal dependency identifiers, and channel scheduling records are generated, including: Obtain a pre-configured list of parallel bearer channels. The list of parallel bearer channels is pre-established according to the target language translation configuration, and the channel identifier and cache status of each parallel bearer channel are recorded. Extract the modal dependency identifiers recorded in the dependency carrier fields of the translated reasoning fragment; Based on the cache state and modal dependency identifier, determine the parallelism of each translated inference segment; The translated inference segments that are in a parallelizable state are placed into the waiting slot of the corresponding parallel bearer channel, and the translated inference segments that are not in a parallelizable state are placed into the delay slot of the corresponding parallel bearer channel. Based on the segment anchor points and reading status corresponding to the slots to be processed and the delay slots, a channel scheduling record is generated.

6. A movie dubbing inference device supporting multimodal parallel processing, characterized in that, include: The modal track-based delivery module is used to acquire the film data to be translated and the target language translation configuration. Based on the presentation sequence of various heterogeneous modal content in the film data to be translated, it generates modal track-based delivery records. The modal track-based delivery records include segment anchors and modal dependency identifiers. Segment anchors are used to define the re-feedable time periods of the translated content on the film timeline, and modal dependency identifiers are used to represent the cross-segment reading relationship between different segment anchors. The channel scheduling record generation module is used to extract the translated inference segments from the modal track-based bearer record according to the segment anchor points, and allocate parallel bearer channels for each translated inference segment based on the modal dependency identifier, thereby generating a channel scheduling record; the channel scheduling record records the segment anchor points corresponding to each parallel bearer channel and their reading status; The parallel bearer execution module is used to drive each parallel bearer channel to read the corresponding translation inference fragments according to the channel scheduling record, so that each parallel bearer channel generates the channel translation result within the recycle period defined by the fragment anchor point, and writes the channel translation result into the recycling cache indexed by the fragment anchor point. The timing backfeed calibration module is used to perform timing backfeed calibration on the channel translation results when the channel translation results belonging to the same segment anchor point in the recycling cache meet the recycling readiness conditions, so as to obtain the movie translation inference results.

7. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the movie translation reasoning method supporting multimodal parallel processing as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, implements the movie translation inference method supporting multimodal parallel processing as claimed in any one of claims 1 to 5.

9. A movie dubbing inference system supporting multimodal parallel processing, characterized in that, This includes movie data access nodes, channel scheduling nodes, parallel bearer nodes, and recycling / feedback nodes; The film data access node is used to acquire the film data to be translated and the target language translation configuration. Based on the presentation sequence of various heterogeneous modal content in the film data to be translated, it generates a modal track-based carrying record and sends the modal track-based carrying record to the channel scheduling node. The modal track-based carrying record contains segment anchors and modal dependency identifiers. The segment anchors are used to define the re-feedable time period of the translated content on the film timeline, and the modal dependency identifiers are used to represent the cross-segment reading relationship between different segment anchors. The channel scheduling node is used to extract the translated inference segments from the modal track-based bearer record based on the segment anchor point, allocate parallel bearer channels for each translated inference segment based on the modal dependency identifier, generate a channel scheduling record, and send the channel scheduling record to the parallel bearer node; the channel scheduling record records the segment anchor point corresponding to each parallel bearer channel and its reading status; The parallel bearer node is used to drive each parallel bearer channel to read the corresponding translation and inference fragments according to the channel scheduling record, so that each parallel bearer channel generates the channel translation result within the recycle period defined by the fragment anchor point, and writes the channel translation result into the recycling cache indexed by the fragment anchor point. The recycling and recharge node is used to perform time-series recharge calibration on the channel translation results when the channel translation results belonging to the same segment anchor point in the recycling cache meet the recycling readiness conditions, so as to obtain the movie translation inference results.

10. A movie dubbing inference method supporting multimodal parallel processing, characterized in that, include: Obtain the movie data to be processed and the translation constraint configuration. Based on the multimodal temporal correlation in the movie data to be processed, generate temporal carrying information. The temporal carrying information is used to limit the processing boundary of the translated reasoning content in the movie temporal sequence and the reading dependency relationship between different processing boundaries. The translation inference bearer object is constructed based on the timing bearer information, and parallel execution guidelines are configured for the translation inference bearer object according to the reading dependency relationship. The parallel execution guidelines are used to indicate the bearer position and reading status of each translation inference bearer object. Following the parallel execution guidelines, multiple parallel processing paths are driven to read the corresponding translation and inference bearer objects, enabling multiple parallel processing paths to generate path translation results under processing boundary constraints, and writing the path translation results into the result recycling area; When the path translation results in the result recovery area meet the recovery judgment conditions used to characterize the recirculation of multiple path results within the same processing boundary, the path translation results are subjected to time-series calibration processing to obtain the movie translation inference results.