A large model-based speech generation method, device and medium

Abstract

The application discloses a large model-based speech generation method and device and medium, and belongs to the technical field of data processing. The method comprises the following steps: dividing target language text information into multiple speech generation units through a text division model; in response to the confidence that the target language text information is divided into multiple speech generation units through the text division model being lower than a first threshold value, dividing the target language text information into multiple speech generation units according to text structure characteristics and parameter distribution characteristics; in response to the speech generation units meeting a multiplexing condition, obtaining a first speech output result according to the speech generation units; in response to the speech generation units not meeting the multiplexing condition, calling a speech generation processing object to obtain a second speech output result, and combining the first speech output result and the second speech output result to generate a target speech result. The application can improve the overall processing efficiency, resource utilization rate and result multiplexing capability under a parameterized speech task.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a speech generation method, apparatus and medium based on a large model. Background Technology

[0002] Currently, in voice interaction scenarios such as intelligent outbound calls, intelligent notifications, customer service broadcasts, and follow-up reminders, business systems typically need to generate target voice content in real time based on user identifiers, business status, amount, time, organization name, address information, or other contextual parameters. Existing systems generally adopt a processing mode of "generating complete speech from complete text at once," which means first replacing the parameters with the final text, and then sending the complete text into the speech synthesis engine to generate the target recording.

[0003] While this approach is simple to implement, its processing granularity is fixed at the level of entire sentences or paragraphs of text. When only local information in the target text changes, the system still needs to re-generate the speech for the entire text. For a large number of parameterized speech tasks with only local differences, this coarse-grained processing method will cause problems such as repeated calculations, repeated generation, low resource utilization, and a decrease in overall response efficiency. Summary of the Invention

[0004] To overcome the aforementioned technical deficiencies, the purpose of this application is to provide a speech generation method, apparatus, and medium based on a large model. The method includes: receiving an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; acquiring a text segmentation model; dividing the target language text information into multiple speech generation units using the text segmentation model; responding to a situation where the confidence level of dividing the target language text information into multiple speech generation units using the text segmentation model is lower than a first threshold, dividing the target language text information into multiple speech generation units based on text structure features and parameter distribution features; and determining whether each speech generation unit meets the reuse conditions using a full matching strategy or a hierarchical matching strategy. In response to the speech generation unit satisfying the reuse condition, a first speech output result is obtained based on the speech generation unit that satisfies the reuse condition; in response to the speech generation unit not satisfying the reuse condition, a second speech output result is obtained by calling a speech generation processing object based on the speech generation unit that does not satisfy the reuse condition. The speech generation processing object includes at least one of a text-corresponding word-level processing object, a phrase-level processing object, and a sentence-level processing object; a target speech result is generated by combining several first speech output results and several second speech output results. This application can improve the overall processing efficiency, resource utilization, and result reuse capability under parameterized speech tasks.

[0005] The specific technical solutions provided in the embodiments of this application are as follows: In a first aspect, this application provides a speech generation method based on a large model, the method comprising: Receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; Obtain the text segmentation model; The target language text information is divided into multiple speech generation units using the text segmentation model. When the confidence level of dividing the target language text information into multiple speech generation units by the text segmentation model is lower than a first threshold, the target language text information is divided into multiple speech generation units according to the text structure features and parameter distribution features. Whether each speech generation unit meets the reuse conditions is determined by a full matching strategy or a hierarchical matching strategy. In response to the speech generation unit satisfying the multiplexing condition, a first speech output result is obtained based on the speech generation unit that satisfies the multiplexing condition; In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The target speech result is generated by combining several first speech output results and several second speech output results.

[0006] Secondly, this application also provides a speech generation apparatus based on a large model, the apparatus comprising: The receiving module is used to receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; The acquisition module is used to acquire the text segmentation model; The segmentation module is used to segment the target language text information into multiple speech generation units using the text segmentation model; in response to a first threshold being lower than the confidence level of segmenting the target language text information into multiple speech generation units using the text segmentation model, the target language text information is then segmented into multiple speech generation units based on text structure features and parameter distribution features. The judgment module is used to determine whether each speech generation unit meets the reuse conditions through a full matching strategy or a hierarchical matching strategy. The output module is configured to obtain a first speech output result based on the speech generation unit that satisfies the multiplexing condition in response to the speech generation unit satisfying the multiplexing condition. In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The generation module is used to generate a target speech result by combining several first speech output results and several second speech output results.

[0007] Thirdly, a speech generation device based on a large model is also provided, including: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the large model-based speech generation method as described in any of the first aspects.

[0008] Fourthly, this application also provides a computer device, the device comprising: A memory, a processor, and a computer program stored in the memory, the processor executing the computer program to implement the steps of the large model-based speech generation method as described in any of the first aspects.

[0009] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the large-model-based speech generation method described in any of the first aspects.

[0010] Sixthly, this application also provides a computer storage medium, the medium comprising: It stores a computer program that, when executed by a processor, implements the steps of the large-model-based speech generation method described in any of the first aspects.

[0011] Compared with the prior art, the method of the technical solution provided in this application includes: receiving an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; obtaining a text segmentation model; dividing the target language text information into multiple speech generation units using the text segmentation model; in response to a situation where the confidence level of dividing the target language text information into multiple speech generation units using the text segmentation model is lower than a first threshold, dividing the target language text information into multiple speech generation units based on text structure features and parameter distribution features; determining whether each speech generation unit meets the reuse condition using a full matching strategy or a hierarchical matching strategy; in response to a speech generation unit meeting the reuse condition, obtaining a first speech output result based on the speech generation unit meeting the reuse condition; in response to a speech generation unit not meeting the reuse condition, calling a speech generation processing object based on the speech generation unit not meeting the reuse condition to obtain a second speech output result, wherein the speech generation processing object includes at least one of a text-corresponding word-level processing object, a phrase-level processing object, and a sentence-level processing object; and generating a target speech result by combining several first speech output results and several second speech output results. This application can improve the overall processing efficiency, resource utilization, and result reuse capability of parameterized speech tasks.

[0012] The technical solution provided in this application can significantly reduce the proportion of repetitive generation in parameterized speech tasks by processing the target speech content into units; by establishing a reuse mechanism for speech generation units, the reuse rate of speech results and overall resource utilization can be improved; by incrementally generating locally changed content, the generation cost and processing latency of batch speech tasks can be reduced; at the same time, this application is applicable to various parameterized speech service scenarios and has strong scalability and versatility. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 The first flowchart of the speech generation method based on a large model provided in Embodiment 1 of this application; Figure 2 The second flowchart of the speech generation method based on a large model provided in Embodiment 2 of this application; Figure 3 This is a flowchart of parameterized speech content unit generation and reuse provided in Embodiment 2 of this application; Figure 4 This is a flowchart of the automatic segmentation and generation of parameterized speech content provided in Embodiment 2 of this application; Figure 5 This is a structural diagram of the speech generation device based on a large model provided in Embodiment 3 of this application; Figure 6 This is an exemplary system provided for Embodiment Seven of this application, which can be used to implement the various embodiments described in this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] It should be noted that, unless the context explicitly requires it, the words "comprising," "including," and similar terms in the entire specification and claims should be interpreted as including rather than exclusive or exhaustive; that is, meaning "including but not limited to."

[0017] Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0018] Example script: Just a reminder that your ${loanInstitutions} bill of ${overdueAmount} will become overdue tomorrow if you don't deal with it today, which may affect your credit score. Please take your repayment seriously. I won't bother you any longer. Goodbye.

[0019] If 1,000 customers have different amounts owed, and the whole sentence synthesis method is used, the whole sentence needs to be processed 1,000 times for speech generation, and a large amount of stable text content will also be generated 1,000 times.

[0020] Therefore, the existing technology has at least the following drawbacks: The speech generation granularity is too coarse, and the results can only be reused under the condition that the complete text is consistent. It is difficult to identify and extract reusable parts from the target speech content, resulting in limited reuse capabilities. The lack of an incremental processing mechanism for locally changed content leads to a large amount of stable content being repeatedly generated. Batch parameterized speech tasks increase generation costs, processing latency, and resource consumption.

[0021] As a starting point for comparison with existing technologies, there are already publicly available solutions that include processing ideas for caching TTS results or optimizing the text-to-speech process. For example, US7043432B2 (Method and system for text-to-speech caching) discloses a text-to-speech caching method, the core of which is to establish a cache mapping between text and speech output; US7979280B2 (Text to speech synthesis) discloses a unit sequence optimization idea for the TTS synthesis process.

[0022] The above scheme can provide inspiration for speech output reuse or speech generation optimization, but based on the publicly available information, it does not clearly elaborate on the overall processing mechanism of "speech generation unit division - reuse judgment - incremental generation - target speech output combination" in parameterized text scenarios.

[0023] Example 1 This application provides a speech generation method based on a large model, such as... Figure 1 As shown, the method includes: Receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; Obtain the text segmentation model; The target language text information is divided into multiple speech generation units using the text segmentation model. When the confidence level of dividing the target language text information into multiple speech generation units by the text segmentation model is lower than a first threshold, the target language text information is divided into multiple speech generation units according to the text structure features and parameter distribution features. Whether each speech generation unit meets the reuse conditions is determined by a full matching strategy or a hierarchical matching strategy. In response to the speech generation unit satisfying the multiplexing condition, a first speech output result is obtained based on the speech generation unit that satisfies the multiplexing condition; In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The target speech result is generated by combining several first speech output results and several second speech output results.

[0024] Specifically, this application relates to intelligent voice interaction, text-to-speech, and voice result management and reuse technologies, and in particular to a method for dividing parameterized text content into voice generation units, reusing voice units, generating missing units, and outputting target voice combinations.

[0025] The purpose of this application is to provide a modular generation and reuse scheme for parameterized speech content. By dividing the target speech content to be output into multiple speech generation units, the speech generation units with reuse value are identified, managed and called, and speech generation units that do not meet the reuse conditions are generated on demand. The target speech output is formed based on multiple speech generation units, thereby improving the overall processing efficiency, resource utilization and result reuse capability under parameterized speech tasks.

[0026] The boundaries of the speech generation unit can be specified manually or automatically identified by the system, thereby enabling the parameterized speech content to be automatically decomposed into variable text units and stable text units.

[0027] The highlights of this application mainly include: The complete speech generation task is abstracted into multiple independently processable speech generation units, thereby expanding the applicability of the solution. Establish a multi-granularity speech result reuse mechanism, not limited to single segment level or whole sentence level implementation; It supports incremental generation only for speech generation units that do not meet the reuse conditions, reducing redundant calculations; It supports the combination of multiple speech generation units to form a target speech result, and continues to manage and reuse the result.

[0028] Taking the sentence "A reminder that your overdue bill of ${loanInstitutions} amounting to ${overdueAmount} will become overdue tomorrow if not addressed today, potentially affecting your credit score. Please take your repayment seriously. Goodbye." as an example, if 1000 customers have different outstanding amounts, existing solutions would require generating the entire sentence 1000 times. This application extracts reusable speech generation units, repeatedly generating only units containing variable information while reusing stable text units. Therefore, the overall character consumption and generation frequency are significantly reduced. Based on current testing of common speech patterns, in typical scenarios with a high proportion of stable text, the number of characters generated for speech generation can typically be reduced by 60%.

[0029] The beneficial effects of the technical solutions provided in this application are: This application can improve the overall processing efficiency, resource utilization, and result reuse capability of parameterized speech tasks.

[0030] The technical solution provided in this application can significantly reduce the proportion of repetitive generation in parameterized speech tasks by processing the target speech content into units; by establishing a reuse mechanism for speech generation units, the reuse rate of speech results and overall resource utilization can be improved; by incrementally generating locally changed content, the generation cost and processing latency of batch speech tasks can be reduced; at the same time, this application is applicable to various parameterized speech service scenarios and has strong scalability and versatility.

[0031] Example 2 This application provides a speech generation method based on a large model, such as... Figure 2 As shown, the method includes: This application addresses at least the following technical problems: How to identify reusable speech generation units from parameterized text content; How to avoid repetitive speech generation when the target speech content changes locally; How to establish a reuse, scheduling, and result management mechanism for speech generation units; How to generate, combine, and output target speech results while ensuring the accuracy of the target speech.

[0032] This application provides a method for the modular generation and reuse of parameterized speech content, which can be deployed in AI outbound calling systems, notification broadcasting systems, intelligent customer service systems, or other processing systems that require speech output of parameterized text.

[0033] Step S01: Receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; Obtain the text segmentation model; The target language text information is divided into multiple speech generation units using the text segmentation model. When the confidence level of dividing the target language text information into multiple speech generation units by the text segmentation model is lower than the first threshold (0.85), the target language text information is divided into multiple speech generation units according to the text structure features and parameter distribution features.

[0034] Specifically, a target speech generation request is received. The request may include a task identifier, a service node identifier, source text content, parameter information, broadcast configuration, and other control information related to speech output. The source text content may be templated text, semi-structured text, or natural language text constrained by a method.

[0035] The system determines the unit division results of the target speech content. Based on explicit tags, text structure features, parameter distribution features, semantic boundaries, punctuation boundaries, pause boundaries, models, or combinations thereof, the system can divide the target text information into multiple speech generation units. These speech generation units can correspond to word-level, phrase-level, sentence-level, or other granular processing objects. In some implementations, the system automatically identifies the text region where parameter placeholder content is located and, combined with sentence structure, punctuation positions, and preset thresholds, decomposes the target text information into variable text units and stable text units.

[0036] Step S011: In response to the presence of an explicit partition marker in the target language text information, the explicit partition marker is used as the language text priority boundary, and the target language text information is divided into multiple speech generation units through the explicit partition marker. In response to the absence of explicit partitioning markers in the target language text information, the position of the parameter placeholder content is obtained, the punctuation marks before and after the parameter and the semantic pause position are determined based on the position of the parameter placeholder content, the language text candidate boundary is determined based on the punctuation marks before and after the parameter and the semantic pause position, and the target language text information is divided into multiple speech generation units through the language text candidate boundary. When the length of the language text between the language text candidate boundaries is less than the second threshold, the language text information of the adjacent language text candidate boundaries is merged, and a speech generation unit is generated based on the merged language text information; the second threshold is about 5 characters. When the target text numerical date and / or organization name and / or proper noun in the target language text information has independent broadcasting meaning, an independent word-level or phrase-level speech generation unit is generated based on the target language text information; The language text information containing parameter placeholders is determined as variable text units, and variable speech generation units are generated based on the variable text units. Language text information that does not contain parameter placeholders and meets the preset length threshold (5 characters) and semantic integrity requirements is identified as stable text units, and stable speech generation units are generated based on the stable text units.

[0037] Specifically, the speech generation unit division can be carried out in the following ways: if there is an explicit division mark in the source text, the explicit division mark is used as the priority boundary; if there is no explicit division mark, the position of the parameter placeholder content is detected, and the candidate boundary is determined by combining the punctuation before and after the parameter, the semantic pause position, and the text length threshold; when the text length between candidate boundaries is less than the preset threshold (5 characters), the text information of adjacent candidate boundaries can be merged to avoid forming excessively short units; when a certain value, date, organization name, or proper noun has independent broadcasting meaning, it can be treated as a word-level or phrase-level speech generation unit. Furthermore, the system can prioritize the identification of text regions containing parameter placeholder content as variable text units, and identify text regions that do not contain parameter placeholder content and meet the length threshold (5 characters) and semantic integrity requirements as stable text units, so as to perform differentiated reuse judgment and on-demand generation in the future.

[0038] Step S012: Obtain the input items of the text segmentation model. The input items include at least one of the following: target language text information, parameter position markers, punctuation features, sentence features, and context features. The target language text information is segmented using the input items of the text segmentation model to obtain candidate boundary positions and unit category labels, wherein the unit category labels are used to distinguish the variable text units, stable text units, and transitional text units. The target language text information is divided into multiple speech generation units based on the candidate boundary positions and unit category labels.

[0039] Specifically, unit partitioning can also be determined by a large model strategy. The large model can be a text partitioning model based on sequence labeling or boundary prediction. The input to the text partitioning model includes original text information, parameter position markers, punctuation features, sentence features, and contextual features. The text partitioning model transforms the original text information and various marked features into mathematical vectors that the model can understand. An adaptive sliding window is used to jointly encode multiple consecutive sentences, modeling the complex dependencies between sentences. Each potential segmentation point (usually the end of a sentence or after a specific token) is scored, and a confidence threshold is set. When the probability that a certain position predicts as a boundary by the text partitioning model exceeds this confidence threshold, that position is considered a boundary. The candidate boundary position is locked. Based on the candidate boundary position predicted by the text segmentation model, the text is divided into preliminary semantic segments. The semantic vector of the segment is input into a fully connected classification layer (Softmax) for multi-class prediction. According to a predefined label system, the unit category probability distribution of the segment is output. Finally, the category with the highest predicted probability is taken as the unit category label of the segment. That is, the output of the text segmentation model is either the candidate boundary position or the unit category label. The model output can be used in conjunction with a preset method. When the model confidence is lower than the first threshold (0.85), it falls back to the method of dividing the target language text information into multiple speech generation units based on text structure features and parameter distribution features. Furthermore, the text segmentation model can also output unit category labels to distinguish variable text units, stable text units, and optional transitional text units. The above description is used to support large model strategy schemes and does not limit the specific model architecture.

[0040] Step S02: Determine whether each speech generation unit meets the reuse conditions using a full matching strategy or a hierarchical matching strategy.

[0041] Specifically, for each speech generation unit, it is determined whether the reuse conditions are met, and the unit result query is performed; the reuse conditions can be determined by the unit text, parameter replacement results, broadcast style, speaker, speech rate, language, sampling rate, service identifier or a combination thereof; if a speech generation unit meets the reuse conditions, the speech result corresponding to the historical speech generation unit is directly obtained.

[0042] Step S021: Determine whether each speech generation unit meets the reuse conditions using a full matching strategy; When the unit parameters of the speech generation unit are all identical to the unit parameters of the historical speech generation units, it is determined that the speech generation unit meets the reuse condition; the unit parameters include the text content of the speech generation unit, the parameter replacement result, and the speech playback configuration parameters, wherein the speech playback configuration parameters include at least one of the following: speaker parameters, timbre parameters, speech rate parameters, language parameters, and sampling rate parameters; or, A hierarchical matching strategy is used to determine whether each speech generation unit meets the reuse conditions; First, match the text content and parameter replacement result of the speech generation unit, and then match the speech broadcast configuration parameters. If any unit parameter of the speech generation unit is inconsistent with the unit parameter of the historical speech generation unit, then it is determined that the speech generation unit does not meet the reuse condition. If the speech generation unit is a variable speech generation unit, then it is determined that the speech generation unit does not meet the reuse condition. If the speech generation unit is a stable speech generation unit, then the speech generation unit is determined to meet the reuse condition.

[0043] Specifically, the reuse condition of a speech generation unit can be determined using a full-match strategy, meaning that the speech generation unit is deemed to meet the reuse condition when all parameters—text content, parameter replacement result, speaker parameters, timbre parameters, speech rate parameters, language parameters, and sampling rate parameters—are consistent. In other implementations, the reuse condition can also employ a layered matching strategy, where the text content and parameter replacement result are matched first, followed by the broadcast configuration parameters that affect acoustic performance. If any parameter of the speech generation unit is inconsistent with the parameters of a historical speech generation unit, the speech generation unit is deemed not to meet the reuse condition. For example, when the speaker is the same but the speech rate is different, it can be determined that the reuse condition is not met to avoid inconsistent output results caused by direct reuse.

[0044] Step S03: In response to the speech generation unit satisfying the multiplexing condition, the first speech output result is obtained based on the speech generation unit that satisfies the multiplexing condition. In response to the speech generation unit not meeting the reuse condition, a second speech output result is obtained by calling a speech generation processing object based on the speech generation unit that does not meet the reuse condition. The speech generation processing object includes at least one of a text-corresponding word-level processing object, a phrase-level processing object, and a sentence-level processing object.

[0045] Specifically, for speech generation units that do not meet the reuse conditions or for which no target speech result is found, the speech generation capability is invoked to generate speech on demand, thereby obtaining the speech output result of the corresponding speech generation unit, and the speech result of the unit can be included in subsequent reuse management.

[0046] Here, on-demand generation means that for a speech generation unit that does not meet the reuse conditions or does not find the target speech result, the speech generation capability is only invoked for that speech generation unit, while other speech generation units that have met the reuse conditions are not generated again, thereby reducing the overall generation overhead.

[0047] Step S031: In response to the speech generation unit's language text information including content with independent broadcasting meaning, the speech generation unit is processed by the corresponding word-level processing object to obtain a second speech output result; the content with independent broadcasting meaning includes at least one of the following: amount, date, organization name, person name, and place name; If the language text information of the speech generation unit includes a text segment consisting of multiple words and having a fixed semantic meaning, then the speech generation unit is processed by the phrase-level processing object to obtain a second speech output result. The speech generation unit responds to the language text information, which includes short sentences, clauses, and stable segments of speech divided by semantic boundaries. The speech generation unit is then processed by the segment-level processing object to obtain the second speech output result.

[0048] Specifically, the speech generation unit may include processing objects at different granularities, including word-level, phrase-level, sentence-level, or other granularities.

[0049] Among them, word-level speech generation units usually correspond to content with relatively independent broadcasting meaning, such as amount, date, organization name, person name, and place name; phrase-level speech generation units usually correspond to text segments composed of multiple words and having fixed semantics as a whole, such as "If it is not processed today" or "Please process it as soon as possible"; sentence-level speech generation units usually correspond to short sentences, clauses, and stable speech segments divided by punctuation, pauses, and semantic boundaries, such as "It will be overdue tomorrow, which may affect your personal credit."

[0050] In practical applications, speech generation units of different granularities can be mixed together in the same target text; for frequently changing parameter content, word-level or phrase-level units can be used to process the object; for frequently used and semantically stable utterance content, sentence-level units can be used to process the object.

[0051] For example, in the sentence "A reminder that your overdue Amount bill from ${loanInstitutions} will become overdue tomorrow if not processed today, which may affect your credit score. Please take your repayment seriously. I won't bother you any longer, goodbye.", "${loanInstitutions}" and "${overdue Amount}" can be processed as word-level or phrase-level variable-related units, "if not processed today" can be processed as a phrase-level unit, and "becomes overdue tomorrow, which may affect your credit score" can be processed as a sentence-level stable text unit. The system can perform speech generation individually on the unmatched speech generation units based on the frequency of change and reuse value of each speech generation unit.

[0052] Step S04: Generate the target speech result by combining several first speech output results and several second speech output results.

[0053] Specifically, the target speech output is formed based on multiple speech generation units. The formation method can be sequential splicing, temporal arrangement, hierarchical combination, synthetic output with pause control, or other combination methods that can produce the target speech result.

[0054] Step S041: In response to a scenario where no additional audio signal correction is required for the speech output scenario, the speech output results corresponding to multiple speech generation units are spliced ​​together according to the original position order of the speech generation units in the target language text information to generate the target speech result. In response to scenarios where additional audio signal correction is required for voice output, the target voice result for the current voice dialogue level is generated by multiple voice generation units corresponding to the currently triggered speech content based on the current voice dialogue level and the current business node; for multiple voice generation units corresponding to speech content at voice dialogue levels that have not been triggered, the generation of the corresponding target voice result is stopped.

[0055] Specifically, the system concatenates the speech results of the corresponding speech generation units according to their original order in the source text to form the current target speech output. In other implementations, the system combines the current speech dialogue level, the current business node, or the current interaction stage to form the target speech output of the current level only for the speech generation units corresponding to the currently triggered dialogue content. For subsequent levels of dialogue content that have not been triggered, no output is generated yet. Unit division, reuse judgment, on-demand generation, and speech result combination are performed after the dialogue progresses to the corresponding speech dialogue level.

[0056] Step S05: Perform fluency processing on the target speech result; The fluency processing of the target speech result includes: The speech results of adjacent speech generation units of the target speech result are subjected to boundary processing, wherein the boundary processing includes at least one of volume normalization, pause length adjustment, boundary fade-in and fade-out processing, silence compensation, prosodic smoothing, and fundamental frequency continuity adjustment.

[0057] Specifically, fluency processing can also be performed during the formation of the target speech output; the fluency processing includes: performing volume normalization, pause length adjustment, boundary fade-in / fade-out processing, silence compensation, prosody smoothing, or fundamental frequency continuity adjustment on the speech results of adjacent speech generation units, so as to reduce the abrupt boundary problems that may be caused by direct combination; for scenarios that do not require additional audio signal correction, the speech results of adjacent speech generation units can also be directly and sequentially spliced ​​to form the target speech output.

[0058] Step S06: Perform result management on the target speech result; The result management of the target speech result includes: A voice retrieval index is built based on the voice script ID, and the target voice results are distinguished based on the voice timbre ID. The target voice result is reusable or incrementally associated based on the voice association item, which includes at least one of the following: voice dialogue level identifier, business node identifier, voice generation unit sequence number, voice generation unit type, and voice broadcast configuration parameters.

[0059] Specifically, result management is performed on the target speech output; result management may include any one or more of the following: whole result storage, index registration, reuse marking, lifecycle control, hit statistics, and subsequent call return.

[0060] The result management is based on an index established by the voice script ID and the voice timbre ID is combined to distinguish the target voice output results. It can also be further associated with at least one of the following: dialogue level identifier, business node identifier, unit number, unit type, and broadcast configuration parameters, for subsequent retrieval, reuse, or incremental generation judgment.

[0061] like Figure 3 , Figure 4 As shown, in this application, the operator inputs a speech script template with variables in the node configuration and indicates the candidate unit boundary by explicit marking. When the front end calls the preview interface, the server first queries whether the complete speech result can be reused. If it does not match, the target text information is processed into speech units, and each speech generation unit is judged to meet the reuse conditions. Speech generation is performed on the speech generation units that do not meet the reuse conditions, and finally the complete speech result is obtained by combining them.

[0062] The system receives a batch of parameterized business data. Each data item corresponds to the same basic template but has different local parameters. When processing each item, the server prioritizes reusing the same stable speech unit, only performs generation on the variable speech generation unit, and further manages the resulting target speech result for direct use by subsequent tasks.

[0063] The system does not rely on manual insertion of markers. Instead, it automatically identifies candidate unit boundaries based on parameter positions, punctuation distribution, semantic structure, pause features, text length thresholds, or trained text strategy models, thereby determining suitable speech generation units for reuse and generation scheduling. As one implementation method, the system first marks the positions before and after the parameter placeholder content as primary candidate boundaries, and then filters secondary candidate boundaries based on punctuation information such as commas, periods, and question marks. If the text length between adjacent candidate boundaries is too short, they are merged. If a text model strategy is used, the boundary prediction model outputs the candidate boundary probability, and candidate boundaries exceeding a preset threshold are retained.

[0064] For content containing relatively independent broadcast meanings, such as amounts, dates, institution names, product names, or place names, the corresponding content can be divided into word-level or phrase-level speech generation units. For example, in the text "Your bill amount at ${institution} on ${date} is ${amount} yuan", the date, institution name, and amount are treated as independent speech units, while the remaining stable wording is treated as sentence-level units, thus forming a mixed-granularity speech unit structure.

[0065] When combining speech to form the target speech output, the system can perform boundary optimization processing on the speech results of adjacent speech units. For example, when there is a significant volume difference between the end of the speech result of the previous speech generation unit and the beginning of the speech result of the next speech generation unit, volume normalization is performed first. When the pause at the unit boundary is insufficient or too long, the silence interval is adjusted. When the boundary transition is abrupt, fade-in / fade-out processing or prosody smoothing processing can be performed to improve the overall listening experience. For scenarios with higher real-time requirements, some boundary optimization steps can be omitted, and the sequentially concatenated speech output result can be directly output.

[0066] The speech generation method provided in this application embodiment can be improved and optimized in several ways without departing from the technical solution of this application, and these improvements and optimizations should also be considered within the scope of protection of this application.

[0067] The core of this application is not limited to using a specific segmentation character, a specific buffer structure, or a specific audio splicing method, but rather lies in: for parameterized speech content, first identifying and dividing it into independently processable speech generation units, then performing multiplexing scheduling or on-demand generation on each speech generation unit according to multiplexing conditions, and finally forming the target speech output. Therefore, any equivalent solution that falls within the above technical concept can be considered an alternative implementation of this application.

[0068] The beneficial effects of the technical solutions provided in this application are: This application can improve the overall processing efficiency, resource utilization, and result reuse capability of parameterized speech tasks.

[0069] The technical solution provided in this application can significantly reduce the proportion of repetitive generation in parameterized speech tasks by processing the target speech content into units; by establishing a reuse mechanism for speech generation units, the reuse rate of speech results and overall resource utilization can be improved; by incrementally generating locally changed content, the generation cost and processing latency of batch speech tasks can be reduced; at the same time, this application is applicable to various parameterized speech service scenarios and has strong scalability and versatility.

[0070] Example 3 This application provides a speech generation device based on a large model, such as... Figure 5 As shown, the device includes: a receiving module, an acquisition module, a division module, a judgment module, an output module, a generation module, and a post-processing module.

[0071] In this embodiment, the receiving module is used to receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; The acquisition module is used to acquire the text segmentation model; The segmentation module is used to segment the target language text information into multiple speech generation units using the text segmentation model; in response to a first threshold being lower than the confidence level of segmenting the target language text information into multiple speech generation units using the text segmentation model, the target language text information is then segmented into multiple speech generation units based on text structure features and parameter distribution features. The judgment module is used to determine whether each speech generation unit meets the reuse conditions through a full matching strategy or a hierarchical matching strategy. The output module is configured to obtain a first speech output result based on the speech generation unit that satisfies the multiplexing condition in response to the speech generation unit satisfying the multiplexing condition. In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The generation module is used to generate a target speech result by combining several first speech output results and several second speech output results.

[0072] In this embodiment, the segmentation module is used to, in response to the presence of explicit segmentation markers in the target language text information, use the explicit segmentation markers as language text priority boundaries and divide the target language text information into multiple speech generation units through the explicit segmentation markers; In response to the absence of explicit partitioning markers in the target language text information, the position of the parameter placeholder content is obtained, the punctuation marks before and after the parameter and the semantic pause position are determined based on the position of the parameter placeholder content, the language text candidate boundary is determined based on the punctuation marks before and after the parameter and the semantic pause position, and the target language text information is divided into multiple speech generation units through the language text candidate boundary. When the length of the language text between the language text candidate boundaries is less than the second threshold, the language text information of the adjacent language text candidate boundaries is merged, and a speech generation unit is generated based on the merged language text information. When the target text numerical date and / or organization name and / or proper noun in the target language text information has independent broadcasting meaning, an independent word-level or phrase-level speech generation unit is generated based on the target language text information; The language text information containing parameter placeholders is determined as variable text units, and variable speech generation units are generated based on the variable text units. Language text information that does not contain parameter placeholders and meets the preset length threshold and semantic integrity requirements is identified as a stable text unit, and a stable speech generation unit is generated based on the stable text unit.

[0073] In one embodiment, the judgment module is used to determine whether each speech generation unit meets the reuse conditions through a full matching strategy or a hierarchical matching strategy, including: The full matching strategy is used to determine whether each speech generation unit meets the reuse conditions; When the unit parameters of the speech generation unit are all identical to the unit parameters of the historical speech generation units, it is determined that the speech generation unit meets the reuse condition; the unit parameters include the text content of the speech generation unit, the parameter replacement result, and the speech playback configuration parameters, wherein the speech playback configuration parameters include at least one of the following: speaker parameters, timbre parameters, speech rate parameters, language parameters, and sampling rate parameters; or, A hierarchical matching strategy is used to determine whether each speech generation unit meets the reuse conditions; First, match the text content and parameter replacement result of the speech generation unit, and then match the speech broadcast configuration parameters. If any unit parameter of the speech generation unit is inconsistent with the unit parameter of the historical speech generation unit, then it is determined that the speech generation unit does not meet the reuse condition. If the speech generation unit is a variable speech generation unit, then it is determined that the speech generation unit does not meet the reuse condition. If the speech generation unit is a stable speech generation unit, then the speech generation unit is determined to meet the reuse condition.

[0074] In one embodiment, the generation module is configured to, in response to a scenario where no additional audio signal correction is required in the speech output scenario, concatenate the speech output results corresponding to multiple speech generation units according to the original position order of the speech generation units in the target language text information to generate a target speech result. In response to scenarios where additional audio signal correction is required for voice output, the target voice result for the current voice dialogue level is generated by multiple voice generation units corresponding to the currently triggered speech content based on the current voice dialogue level and the current business node; for multiple voice generation units corresponding to speech content at voice dialogue levels that have not been triggered, the generation of the corresponding target voice result is stopped.

[0075] It is understood that this application provides a processing method for dividing parameterized text content into speech generation units; a reuse judgment and result scheduling mechanism for speech generation units; an incremental speech generation mechanism that only performs on-demand generation for speech generation units that do not meet the reuse conditions; as shown in Table 1 below: a processing mechanism for forming target speech results based on multiple speech generation units and continuing to manage the results; and an implementation method for determining candidate boundaries of speech generation units through explicit methods, automatic methods, or model strategies.

[0076] Table 1

[0077] The speech generation unit can be a word-level, phrase-level, sentence-level, or hierarchical combination unit, and is not limited to fixed-length segmentation; the management method of the reused results can adopt caching, indexing, object storage, database records, or other searchable result management mechanisms; the formation method of the target speech result can be direct splicing, temporal arrangement, hierarchical mixing, or other processing methods that can achieve combined output; in addition to financial outbound calls, this application can also be applied to parameterized speech business scenarios such as notification reminders, marketing outbound calls, customer service broadcasts, intelligent follow-up visits, navigation broadcasts, and voice assistants.

[0078] The beneficial effects of the technical solutions provided in this application are: This application can improve the overall processing efficiency, resource utilization, and result reuse capability of parameterized speech tasks.

[0079] The technical solution provided in this application can significantly reduce the proportion of repetitive generation in parameterized speech tasks by processing the target speech content into units; by establishing a reuse mechanism for speech generation units, the reuse rate of speech results and overall resource utilization can be improved; by incrementally generating locally changed content, the generation cost and processing latency of batch speech tasks can be reduced; at the same time, this application is applicable to various parameterized speech service scenarios and has strong scalability and versatility.

[0080] Example 4 This application also provides a speech generation device based on a large model, comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the following large-model-based speech generation method: Receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; Obtain the text segmentation model; The target language text information is divided into multiple speech generation units using the text segmentation model. When the confidence level of dividing the target language text information into multiple speech generation units by the text segmentation model is lower than a first threshold, the target language text information is divided into multiple speech generation units according to the text structure features and parameter distribution features. Whether each speech generation unit meets the reuse conditions is determined by a full matching strategy or a hierarchical matching strategy. In response to the speech generation unit satisfying the multiplexing condition, a first speech output result is obtained based on the speech generation unit that satisfies the multiplexing condition; In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The target speech result is generated by combining several first speech output results and several second speech output results.

[0081] Example 5 This application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can perform the following method for speech generation based on a large model: Receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; Obtain the text segmentation model; The target language text information is divided into multiple speech generation units using the text segmentation model. When the confidence level of dividing the target language text information into multiple speech generation units by the text segmentation model is lower than a first threshold, the target language text information is divided into multiple speech generation units according to the text structure features and parameter distribution features. Whether each speech generation unit meets the reuse conditions is determined by a full matching strategy or a hierarchical matching strategy. In response to the speech generation unit satisfying the multiplexing condition, a first speech output result is obtained based on the speech generation unit that satisfies the multiplexing condition; In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The target speech result is generated by combining several first speech output results and several second speech output results.

[0082] The beneficial effects of the technical solutions provided in this application are: This application can improve the overall processing efficiency, resource utilization, and result reuse capability of parameterized speech tasks.

[0083] The technical solution provided in this application can significantly reduce the proportion of repetitive generation in parameterized speech tasks by processing the target speech content into units; by establishing a reuse mechanism for speech generation units, the reuse rate of speech results and overall resource utilization can be improved; by incrementally generating locally changed content, the generation cost and processing latency of batch speech tasks can be reduced; at the same time, this application is applicable to various parameterized speech service scenarios and has strong scalability and versatility.

[0084] Example 6 This application also provides a computer program product, including a computer program that, when executed by a processor, can implement the following method for speech generation based on a large model: Receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; Obtain the text segmentation model; The target language text information is divided into multiple speech generation units using the text segmentation model. When the confidence level of dividing the target language text information into multiple speech generation units by the text segmentation model is lower than a first threshold, the target language text information is divided into multiple speech generation units according to the text structure features and parameter distribution features. Whether each speech generation unit meets the reuse conditions is determined by a full matching strategy or a hierarchical matching strategy. In response to the speech generation unit satisfying the multiplexing condition, a first speech output result is obtained based on the speech generation unit that satisfies the multiplexing condition; In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The target speech result is generated by combining several first speech output results and several second speech output results.

[0085] The beneficial effects of the technical solutions provided in this application are: This application can improve the overall processing efficiency, resource utilization, and result reuse capability of parameterized speech tasks.

[0086] The technical solution provided in this application can significantly reduce the proportion of repetitive generation in parameterized speech tasks by processing the target speech content into units; by establishing a reuse mechanism for speech generation units, the reuse rate of speech results and overall resource utilization can be improved; by incrementally generating locally changed content, the generation cost and processing latency of batch speech tasks can be reduced; at the same time, this application is applicable to various parameterized speech service scenarios and has strong scalability and versatility.

[0087] Example 7 This application provides a computer storage medium, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it performs the following steps: Receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; Obtain the text segmentation model; The target language text information is divided into multiple speech generation units using the text segmentation model. When the confidence level of dividing the target language text information into multiple speech generation units by the text segmentation model is lower than a first threshold, the target language text information is divided into multiple speech generation units according to the text structure features and parameter distribution features. Whether each speech generation unit meets the reuse conditions is determined by a full matching strategy or a hierarchical matching strategy. In response to the speech generation unit satisfying the multiplexing condition, a first speech output result is obtained based on the speech generation unit that satisfies the multiplexing condition; In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The target speech result is generated by combining several first speech output results and several second speech output results.

[0088] This application can improve the overall processing efficiency, resource utilization, and result reuse capability of parameterized speech tasks.

[0089] The technical solution provided in this application can significantly reduce the proportion of repetitive generation in parameterized speech tasks by processing the target speech content into units; by establishing a reuse mechanism for speech generation units, the reuse rate of speech results and overall resource utilization can be improved; by incrementally generating locally changed content, the generation cost and processing latency of batch speech tasks can be reduced; at the same time, this application is applicable to various parameterized speech service scenarios and has strong scalability and versatility.

[0090] Figure 6 This is an exemplary system provided for Embodiment Seven of this application, which can be used to implement the various embodiments described in this application; like Figure 6 As shown, in some embodiments, the system can function as any of the aforementioned devices for large-model-based speech generation in the various embodiments. In some embodiments, the system may include one or more computer-readable media (e.g., system memory or NVM / storage device) having a result and one or more processors (e.g., one or more processors) coupled to the one or more computer-readable media and configured to execute the result to implement the module thereby performing the actions described in this application.

[0091] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0092] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A speech generation method based on a large model, characterized in that, The method includes: Receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; Obtain the text segmentation model; The target language text information is divided into multiple speech generation units using the text segmentation model. When the confidence level of dividing the target language text information into multiple speech generation units by the text segmentation model is lower than a first threshold, the target language text information is divided into multiple speech generation units according to the text structure features and parameter distribution features. A hierarchical matching strategy is used to determine whether each speech generation unit meets the reuse conditions; In response to the speech generation unit satisfying the multiplexing condition, a first speech output result is obtained based on the speech generation unit that satisfies the multiplexing condition; In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The target speech result is generated by combining several first speech output results and several second speech output results; Based on the current voice dialogue level, current business node, or current interaction stage, only the voice generation unit corresponding to the currently triggered dialogue content will generate the target voice output of the current level. For subsequent level dialogue content that has not been triggered, it will not be generated for the time being. Unit division, reuse judgment, on-demand generation, and voice result combination will be performed after the dialogue progresses to the corresponding voice dialogue level. A hierarchical matching strategy is used to determine whether each speech generation unit meets the reuse conditions, including: First, match the text content and parameter replacement result of the speech generation unit, and then match the speech broadcast configuration parameters. If any unit parameter of the speech generation unit is inconsistent with the unit parameter of the historical speech generation unit, then it is determined that the speech generation unit does not meet the reuse condition. If the speech generation unit is a variable speech generation unit, then it is determined that the speech generation unit does not meet the reuse condition. If the speech generation unit is a stable speech generation unit, then the speech generation unit is determined to meet the reuse condition.

2. The speech generation method based on a large model according to claim 1, characterized in that, The process of dividing the target language text information into multiple speech generation units based on text structure features and parameter distribution features includes: In response to the presence of explicit partition markers in the target language text information, the explicit partition markers are used as language text priority boundaries, and the target language text information is divided into multiple speech generation units through the explicit partition markers. In response to the absence of explicit partitioning markers in the target language text information, the position of the parameter placeholder content is obtained, the punctuation marks before and after the parameter and the semantic pause position are determined based on the position of the parameter placeholder content, the language text candidate boundary is determined based on the punctuation marks before and after the parameter and the semantic pause position, and the target language text information is divided into multiple speech generation units through the language text candidate boundary. When the length of the language text between the language text candidate boundaries is less than the second threshold, the language text information of the adjacent language text candidate boundaries is merged, and a speech generation unit is generated based on the merged language text information. When the target text numerical date and / or organization name and / or proper noun in the target language text information has independent broadcasting meaning, an independent word-level or phrase-level speech generation unit is generated based on the target language text information; The language text information containing parameter placeholders is determined as variable text units, and variable speech generation units are generated based on the variable text units. Language text information that does not contain parameter placeholders and meets the preset length threshold and semantic integrity requirements is identified as a stable text unit, and a stable speech generation unit is generated based on the stable text unit.

3. The speech generation method based on a large model according to claim 2, characterized in that, The process of dividing the target language text information into multiple speech generation units using the text segmentation model includes: Obtain the input items of the text segmentation model, wherein the input items include at least one of the target language text information, parameter position markers, punctuation features, sentence features, and context features; The target language text information is segmented using the input items of the text segmentation model to obtain candidate boundary positions and unit category labels, wherein the unit category labels are used to distinguish the variable text units, stable text units, and transitional text units. The target language text information is divided into multiple speech generation units based on the candidate boundary positions and unit category labels.

4. The speech generation method based on a large model according to claim 1, characterized in that, The step of calling the speech generation processing object based on the speech generation unit that does not meet the reuse conditions to obtain the second speech output result includes: In response to the fact that the language text information of the speech generation unit includes content with independent broadcasting meaning, the speech generation unit is processed by the word-level processing object corresponding to the text to obtain a second speech output result; the content with independent broadcasting meaning includes at least one of the following: amount, date, organization name, person name, and place name; In response to the speech generation unit's language text information including a text segment composed of multiple words and having a fixed semantic meaning, the speech generation unit is processed by the phrase-level processing object to obtain a second speech output result. In response to the language text information of the speech generation unit, which includes short sentences, clauses, and stable segments of speech divided by semantic boundaries, the speech generation unit is processed by the segment-level processing object to obtain a second speech output result.

5. The speech generation method based on a large model according to claim 1, characterized in that, The step of generating the target speech result by combining several first speech output results and several second speech output results includes: In response to scenarios where no additional audio signal correction is required for the speech output, the speech output results corresponding to multiple speech generation units are concatenated according to the original position order of the speech generation units in the target language text information to generate the target speech result. In response to scenarios where additional audio signal correction is required for voice output, the target voice result for the current voice dialogue level is generated by multiple voice generation units corresponding to the currently triggered speech content based on the current voice dialogue level and the current business node; for multiple voice generation units corresponding to speech content at voice dialogue levels that have not been triggered, the generation of the corresponding target voice result is stopped.

6. The speech generation method based on a large model according to claim 1, characterized in that, After generating the target speech result by combining several first speech output results and several second speech output results, the process includes: The target speech result is then processed for fluency. The fluency processing of the target speech result includes: The speech results of adjacent speech generation units of the target speech result are subjected to boundary processing, wherein the boundary processing includes at least one of volume normalization, pause length adjustment, boundary fade-in and fade-out processing, silence compensation, prosodic smoothing, and fundamental frequency continuity adjustment.

7. The speech generation method based on a large model according to claim 1, characterized in that, After generating the target speech result by combining several first speech output results and several second speech output results, the method further includes: Perform result management on the target speech results; The result management of the target speech result includes: A voice retrieval index is built based on the voice script ID, and the target voice results are distinguished based on the voice timbre ID. The target voice result is reusable or incrementally associated based on the voice association item, which includes at least one of the following: voice dialogue level identifier, business node identifier, voice generation unit sequence number, voice generation unit type, and voice broadcast configuration parameters.

8. A large-model-based speech generation apparatus for implementing the method as described in claim 1, characterized in that, The device includes: The receiving module is used to receive an instruction to generate a target speech output request based on control information, wherein the control information includes target language text information; The acquisition module is used to acquire the text segmentation model; The segmentation module is used to segment the target language text information into multiple speech generation units using the text segmentation model; in response to a first threshold being lower than the confidence level of segmenting the target language text information into multiple speech generation units using the text segmentation model, the target language text information is then segmented into multiple speech generation units based on text structure features and parameter distribution features. The judgment module is used to determine whether each speech generation unit meets the reuse conditions through a full matching strategy or a hierarchical matching strategy. The output module is configured to obtain a first speech output result based on the speech generation unit that satisfies the multiplexing condition in response to the speech generation unit satisfying the multiplexing condition. In response to the speech generation unit not meeting the reuse condition, the speech generation processing object is called according to the speech generation unit that does not meet the reuse condition to obtain the second speech output result. The speech generation processing object includes at least one of the following: text corresponding word-level processing object, phrase-level processing object, and sentence-level processing object. The generation module is used to generate a target speech result by combining several first speech output results and several second speech output results.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the speech generation method based on a large model as described in any one of claims 1 to 7.

Images