Offline large-model intelligent bidirectional translation system oriented to navigation communication guarantee

Abstract

The invention relates to the technical field of ship equipment, discloses an off-line large-model intelligent bidirectional translation system oriented to navigation communication guarantee, and aims to solve the problem that navigation communication among different languages depends on networking requirements of translation software or depends on subjective judgment of accompanying translators in the prior art. According to the technical scheme, the communication content is extracted and translated by using the pre-established model, the subjectivity of manual translation is avoided, networking is not needed, and communication among different languages is realized. According to the invention, the effective voice is extracted, so that the influence of environmental noise is reduced; the voice is converted into the text, communication content can be visually displayed, and information omission is avoided. And the translation unit loaded with the text matching model integrated with the special term of the marine communication can translate special words and sentences in the marine communication, so that the translation accuracy is improved. And the audio recording unit and the storage unit can store communication contents and translation contents, so that subsequent redisk is facilitated.

Description

Technical Field

[0001] This invention relates to the field of marine equipment technology, and in particular to a translation system. Background Technology

[0002] Maritime communication is a core support for ensuring ship navigation safety, operational coordination, and mission execution. Communication between ship formations is even more directly related to operational command, formation coordination, and emergency response efficiency. Currently, maritime communication mainly relies on equipment such as VHF (Very High Frequency) radiotelephones, with real-time voice interaction as the primary method. With the globalization of shipping and the diversification of maritime missions, cross-border crews and diverse communication languages ​​have become the norm, and language barriers have become a key bottleneck in maritime communication. In routine scenarios such as collision avoidance, port scheduling, and formation coordination, language barriers can easily lead to delays in instruction transmission; in emergency scenarios such as Mayday distress calls and engine room malfunctions, language misunderstandings can even lead to catastrophic consequences.

[0003] Current maritime communication technology primarily relies on onboard translators or portable translation software. However, maritime communications often contain specialized terminology such as navigation (e.g., starboard, course correction) and meteorological (e.g., strong winds, typhoon warnings), which translators struggle to accurately interpret. Furthermore, the translations are subject to the translator's subjective judgment, making it difficult to assign responsibility in the event of an accident due to inaccurate translation. Using conventional translation software, on the other hand, depends on cloud servers, but maritime navigation often occurs in satellite signal dead zones or without network coverage, rendering the translation software completely ineffective.

[0004] Therefore, it is evident that avoiding inaccuracies and waste of human resources caused by human translation in maritime communication, and avoiding the networking requirements of software translation, are problems that urgently need to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the problems in current technologies where maritime communication between different languages ​​relies on translation software and requires network connectivity, or on the subjective judgment of accompanying translators. Therefore, this invention provides an offline large-scale intelligent two-way translation system for maritime communication support, which avoids the subjective influence of human translation and the rigid requirement of network connectivity, and enables communication and translation between different languages.

[0006] To address the aforementioned technical problems, this invention provides an offline large-scale intelligent two-way translation system for maritime communication support, comprising:

[0007] Audio processing unit, speech transcription unit, and translation unit;

[0008] The audio processing unit is used to collect ambient audio signals and extract valid speech from the ambient audio signals after receiving a start signal;

[0009] The speech transcription unit converts the valid speech received from the audio processing unit into first language text based on a pre-loaded speech conversion model.

[0010] The translation unit receives the first language text and calls the text matching model to translate the first language text into the second language text; the text matching model is a model that incorporates HNA Communications' proprietary terminology.

[0011] Preferably, the translation unit receives the first language text and invokes a text matching model to translate the first language text into second language text, including:

[0012] The translation unit performs text retrieval by calling a local knowledge base based on the first language text.

[0013] If the similarity between the recorded text in the local knowledge base and the first language text exceeds a preset threshold, the translated text that matches the recorded text will be used as the second language text.

[0014] If the recorded text is not found, a pre-built translation model is invoked to translate the second language text based on the vocabulary information in the first language text.

[0015] Preferably, it also includes: a suggested unit:

[0016] The suggestion providing unit is used to call a pre-built conventional suggestion response matching model after receiving the second language text, and generate response text based on the second language text.

[0017] Preferably, it also includes: a speech generation unit;

[0018] The speech generation unit generates speech based on the second language text and the reply text, and then broadcasts it.

[0019] Preferably, the audio processing unit is configured to collect ambient audio signals and extract valid speech from the ambient audio signals after receiving a start signal, including:

[0020] The audio processing unit collects the ambient audio signal at a preset frequency as the sampling rate and simultaneously calculates the root mean square energy value of each frame of the ambient audio signal.

[0021] The energy value is compared with the set noise threshold. When the energy value of the environmental audio signal in N consecutive frames is greater than the noise threshold, it is determined to be the valid speech and the data buffer is started.

[0022] When the energy value of the environmental audio signal in M ​​consecutive frames is not greater than the noise threshold, it is determined to be silent or noise and data buffering is stopped.

[0023] Both N and M are constants greater than 1.

[0024] Preferably, it also includes: a processing unit and a human-computer interaction unit with display function;

[0025] The processing unit is used to send the start signal to the audio processing unit after receiving the trigger command fed back by the human-computer interaction unit;

[0026] The human-computer interaction unit is also used to receive and display the first language text and the second language text.

[0027] Preferably, the processing unit is further configured to, upon receiving a storage signal, update and save the recorded text that is not found in the local knowledge base to the local knowledge base.

[0028] Preferably, it also includes: an audio recording unit and a storage unit;

[0029] The audio recording unit is used to perform the recording action of the valid voice after receiving the recording instruction sent by the processing unit, and save the recording content to the storage unit;

[0030] The storage unit is used to store the recorded content locally, combining the recording timestamp, the first language text, the second language text, and the structured format of the translation source of the second language text.

[0031] Preferably, the processing unit is further configured to execute an emergency communication mode after receiving a mode switching signal;

[0032] The translation unit in the emergency communication mode is used to call a standard language library and translate the first language text into a second language text; the standard language library is a pre-built database that stores commonly used phrases for emergency maritime communication.

[0033] Preferably, the processing unit is further configured to send a language switching instruction to the translation unit;

[0034] The translation unit is used to, after receiving the first language text, call the corresponding language text matching model according to the language switching instruction to translate the first language text into the second language text.

[0035] This invention provides an offline large-scale intelligent two-way translation system for maritime communication support. Compared to current technologies that rely on translation software for inter-language maritime communication, which requires network connectivity, or on accompanying translators, which involves subjective judgment, this technical solution utilizes a pre-built model to extract and translate communication content, avoiding the subjectivity of manual translation and enabling inter-language communication without network access. In this application, an audio processing unit extracts speech, reducing the impact of environmental noise. A speech transcription unit with a speech conversion model translates the extracted valid speech into first-language text. Converting language to text facilitates the intuitive presentation of communication content and avoids the omission of vocabulary and other information. A translation unit with a text matching model translates the first-language text into easily understandable second-language text, achieving language conversion. Furthermore, the text matching model incorporates maritime communication terminology, enabling the translation of specialized phrases in maritime communication, improving translation accuracy and avoiding misunderstandings of important information. Furthermore, it is recommended that the unit be able to generate response text based on second language text, facilitating crew operations and improving response efficiency. The speech generation unit should be able to play back the second language text and response text aloud, making information retrieval easier for the entire crew. The audio recording and storage units should be able to save communication and translation content for later review. Attached Figure Description

[0036] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0037] Figure 1 This is a structural diagram of an offline large-scale intelligent two-way translation system for maritime communication support provided in an embodiment of the present invention;

[0038] Figure 2 This is an example of the user interface diagram of an offline large-scale intelligent two-way translation system for maritime communication support provided by an embodiment of the present invention.

[0039] The attached figures are labeled as follows: 1 is the audio processing unit, 2 is the speech transcription unit, 3 is the translation unit, 4 is the suggestion providing unit, 5 is the speech generation unit, 6 is the processing unit, and 7 is the human-computer interaction unit. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0041] The core of this invention is to provide an offline large-scale intelligent two-way translation system for maritime communication support, which avoids the subjective influence of human translation and the rigid requirement of network connectivity, and realizes communication and translation between different languages.

[0042] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Figure 1 This is a structural diagram of an offline large-scale intelligent two-way translation system for maritime communication support provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the system includes:

[0044] Audio processing unit 1, speech transcription unit 2, and translation unit 3;

[0045] The audio processing unit 1 is used to collect ambient audio signals and extract valid speech from the ambient audio signals after receiving the start signal;

[0046] The speech transcription unit 2 converts the valid speech received from the audio processing unit 1 into first language text based on a pre-loaded speech conversion model.

[0047] Translation unit 3 receives the first language text and calls the text matching model to translate the first language text into the second language text; the text matching model is a model that incorporates HNA Communications' proprietary terminology.

[0048] First, it should be noted that the offline large-scale intelligent two-way translation system for maritime communication support provided in this application is applied to maritime communication that uses real-time voice interaction as the means of communication. As a translation system, it can be assembled independently into a style that is easy for crew members to carry with them, or it can be integrated into the main control system on the ship to achieve the difference between personal or ship use.

[0049] In this embodiment, the audio processing unit 1 serves as a tool for acquiring communication content and is a pre-processing unit for communication translation. The audio processing unit 1 collects environmental audio signals and extracts valid speech from them. In specific implementations, the audio processing unit 1 operates upon receiving a start signal, which can be a power-on signal. After power is turned on, it continuously collects environmental audio signals. Alternatively, it can be equipped with a corresponding processing unit 6, which operates by receiving instructions from the processing unit 6. Furthermore, the processing unit 6 is also equipped with a human-computer interaction unit 7. The processing unit 6 sends instructions based on feedback from the human-computer interaction unit 7, enabling technicians to actively control the translation process. This improves work efficiency while allowing for selection and control of the translation process and translation mode.

[0050] As can be seen, the audio processing unit 1 in this application simultaneously possesses voice acquisition and voice extraction functions. This embodiment provides a specific feasible solution. In this embodiment, the audio processing unit 1 includes a microphone and a Voice Activity Detection (VAD) module. In a specific implementation, the microphone can be a 4-microphone array, using a Knowles SPH0641 LU4H-1 high signal-to-noise ratio microphone, and supporting beamforming noise reduction algorithms, which can suppress background noise below 85dB, effectively mitigating the problem of high mechanical noise and severe ship vibration affecting sound collection in the maritime communication environment. Furthermore, when collecting environmental audio signals, a preset frequency, such as 16kHz, is used as the sampling rate to collect environmental audio signals, and 16-bit deep PCM format audio signals are acquired, with a data transmission rate of 256kbps.

[0051] The VAD module is used to extract valid speech. Specifically, it can integrate the web RTC VAD library. It synchronously calculates the root mean square energy value of each frame of environmental audio signal collected, compares the energy value with the set noise threshold, and determines that the speech is valid and starts data buffering when the energy value of the environmental audio signal of N consecutive frames is greater than the noise threshold. When the energy value of the environmental audio signal of M consecutive frames is not greater than the noise threshold, it is determined to be silence or noise and data buffering is stopped. N and M are both constants greater than 1.

[0052] When equipped with processing unit 6, audio processing unit 1 can communicate with processing unit 6 via PCIe 4.0 bus. The noise threshold can be dynamically set according to different environments. For example, when processing unit 6 starts up through audio processing unit 1, it silently samples and calibrates the ambient noise energy for 3 seconds, setting the noise threshold to 1.8-2.5 times the noise energy. When comparing, for example, if the energy value exceeds the noise threshold for 3 consecutive frames (20ms per frame), it is determined to be valid speech. If the energy value is below the noise threshold for 8 consecutive frames, it is determined to be silent or noise and buffering is stopped. This ensures that only valid speech data blocks are transmitted to speech transcription unit 2, reducing data transmission volume and interference, and improving translation efficiency.

[0053] The speech transcription unit 2 can transmit data with the audio processing unit 1 through a memory sharing mechanism to receive effective speech data blocks. The speech transcription unit 2 is used to convert effective speech into first language text. It relies on the built-in offline deployed Transformer architecture speech conversion model, which can be the faster-whisper-small model.

[0054] Translation unit 3 communicates with speech transcription unit 2, receives first language text, and invokes a text matching model to translate the first language text into second language text. This embodiment provides a specific operating scenario for translation unit 3. The text matching model loaded by translation unit 3 includes a local knowledge base and a translation model. The local knowledge base is a repository formed by summarizing communication terms through various channels such as historical communication reports and network searches. The translation model is an algorithm model that matches vocabulary information with the translation content. In specific implementation, translation unit 3 receives first language text and invokes the text matching model to translate the first language text into second language text. This includes: translation unit 3 invoking the local knowledge base to perform text retrieval based on the first language text; if the similarity between the retrieved text in the local knowledge base and the first language text exceeds a preset threshold, the translated text matching the retrieved text is used as the second language text; if no retrieved text is found, the pre-built translation model is invoked to translate the second language text based on the vocabulary information in the first language text.

[0055] As can be seen, translation unit 3 divides the translation process of the first language text into two layers. The first is a rough search based on the local knowledge base, which searches for similar sentences to the first language text and uses them as the communication content, with their corresponding translations used as the second language text. If no similar sentences are found, the translation model translates the first language text word by word and sentence by sentence. It is understandable that using the local knowledge base for translation allows for faster translation of communication content compared to using a translation model. Furthermore, the text matching model incorporates HNA's proprietary terminology, thus enabling more accurate translation of the communication content.

[0056] The local knowledge base can be built by scanning external files. For files from different channels, it supports parsing external terminology files in four formats: CSV, PDF, DOCX, and XLSX. Specifically, it uses the pandas library to parse CSV / XLSX tabular data, the python-docx library to extract DOCX text content, and PyPDF2 combined with Tesseract OCR to process PDF scans. After extracting the correspondence between "source language terms and target language terms", it uses difflib. Sequence Matcher to calculate the duplication rate, removes redundant entries with a duplication rate of ≥90%, and stores the data in JSON format to form the local knowledge base. When using the local knowledge base for retrieval, the W Ratio algorithm from the fuzz library is used to calculate the similarity between the first-language text and the source language terms in the local knowledge base. A preset threshold of 90% is used. If the first-language text length is ≤60 characters and the highest similarity is ≥90%, it is considered a "successful match." If the first-language text length is >60 characters, it is divided into clauses using commas, periods, and semicolons. A "successful retrieval" is achieved when all clauses have a similarity of ≥90%; otherwise, a "failed retrieval" is considered. During the retrieval process, the search results can be sorted, and the text with the highest similarity is selected as the output. In cases of failed retrieval, the Helsinki-NLP / opus-mt-en-zh model is used to generate the second-language text.

[0057] In practice, for the translated second language text, an error correction model can be added to process the second language text. For example, the BERT error correction model, trained on a large amount of manual error correction data in the maritime field, can be used to correct mistranslations of terms and errors in grammar and word order in the second language text, effectively improving the translation accuracy.

[0058] This invention provides an offline large-scale intelligent two-way translation system for maritime communication support. Compared to current technologies that rely on translation software for inter-language maritime communication, which requires network connectivity, or on accompanying translators, which involves subjective judgment, this technical solution utilizes a pre-built model to extract and translate communication content, avoiding the subjectivity of manual translation. Furthermore, it eliminates the need for a network connection, enabling communication between different languages. In this application, an audio processing unit extracts speech, reducing the impact of environmental noise. A speech transcription unit equipped with a speech conversion model translates the extracted valid speech into first-language text. This language-to-text conversion facilitates the intuitive presentation of communication content and avoids the omission of vocabulary and other information. A translation unit equipped with a text matching model translates the first-language text into easily understandable second-language text, achieving language conversion. Moreover, the text matching model incorporates maritime communication terminology, enabling the translation of specialized phrases in maritime communication, improving translation accuracy and avoiding misunderstandings of important information.

[0059] Based on the above embodiments, the offline large-scale intelligent bidirectional translation system for maritime communication support provided in this embodiment also includes a processing unit 6 and a human-computer interaction unit 7 with display function. The processing unit 6, upon receiving a trigger command from the human-computer interaction unit 7, sends a start signal to the audio processing unit 1. The human-computer interaction unit 7 is also used to receive and display text in the first language and text in the second language. In specific applications, the processing unit 6 can be an integrated AMD Ryzen 7 series processor (preferably Ryzen 7 8845HS) or an Intel processor with equivalent performance (such as Intel Core i9-13900H). The Ryzen 7 8845HS is an 8-core, 16-thread design with a boost frequency of up to 5.1GHz and integrated Radeon 780M graphics (12CU, 2.7GHz), providing sufficient computing power for parallel inference of AI models. The Intel Core i9-13900H is a 24-core, 32-thread (8P+16E) system with a boost frequency of 5.4GHz and integrated Iris Xe graphics (32EU), optimized for deep learning frameworks under the x86 architecture. The human-machine interaction unit 7 can be an industrial-grade touch screen with a size of no less than 7 inches (resolution 2048×1536, brightness 350cd / m², supports glove touch). The bezel can also integrate a red physical "emergency" button (15mm in diameter) and a "recording start / stop" button. It is configured to realize communication control, bilingual text presentation, emergency triggering, recording management and record archiving through a graphical user interface (GUI). It supports switching between two time-efficiency levels: "quick translation" and "accurate translation". Based on the x86 architecture, it optimizes multi-threaded task scheduling to ensure low latency when the recording, transcription and translation modules run in parallel.

[0060] In addition, the processing unit 6 can communicate with the speech transcription unit 2 and the translation unit 3. The speech transcription unit 2 adopts the x86 architecture based on the core processor and uses a "CPU + integrated graphics" hybrid inference mode. The Radeon 780M integrated graphics of the AMD Ryzen 7 8845HS is responsible for parallel computing tasks such as Mel spectrogram conversion, while the CPU is responsible for decoder inference, with an inference latency of ≤300ms. The Intel Core i9-13900H accelerates feature extraction through the Iris Xe integrated graphics, while the CPU handles the decoding process. The speech transcription unit 2 outputs the first language text and a confidence score (0-1.0). When the confidence score is less than 0.6, "[Pending confirmation]" is marked after the first language text. This text, along with the second language text translated by the translation unit 3, is transmitted to the processing unit 6 and then displayed through the human-computer interaction unit 7. The processing unit 6 is also used to update and save the recorded text that is not found in the local knowledge base to the local knowledge base when a storage signal is received. Specifically, the human-computer interaction unit 7 can also feed back the corresponding information to the processing unit 6 so as to manually add, delete, and modify terms in the local knowledge base for the first language text and the second language text.

[0061] To address emergency communication needs during navigation and to simplify operations, the offline large-scale intelligent two-way translation system for maritime communication support may further include: a suggestion providing unit 4: This unit, upon receiving second language text, invokes a pre-built conventional suggestion-response matching model to generate a response text based on the second language text. Furthermore, it may also include: a speech generation unit 5: This unit generates and broadcasts speech based on the second language text and the response text.

[0062] In this embodiment, the suggestion providing unit 4 communicates with the translation unit 3. The core conventional suggestion response matching model is a 4-bit quantized DeepSeek-R1-Distill-Qwen-1.5B-GGUF model. After inputting second language text, a preset prompt word template is used, such as "You are a marine communication assistant for warships. Provide 3 concise, professional replies to: '{}' (max 50 chars each, match the communication scenario)", to generate 3-5 suggestion responses that conform to maritime communication conventions (such as confirmation type "Confirm the navigation instruction?", inquiry type "Need further detail?"). This information can also be synchronously transmitted to the human-computer interaction unit 7 through the processing unit 6 and presented in the form of touch buttons.

[0063] The speech generation unit 5 communicates with the suggestion providing unit 4 and the processing unit 6, generating and broadcasting speech based on the second language text and response text. Specifically, the speech generation unit 5 may include an audio encoding module and an output module; the audio encoding module uses the Coqui TTS engine, supports 22050Hz sampling rate, 16-bit depth WAV format audio generation, and the speech rate can be adjusted between 0.8 and 1.2 times; the output module uses a 2W full-range speaker (frequency response range 200Hz-10kHz) and is equipped with an audio amplification chip (MAX98357A) to ensure clear and loud broadcast sound.

[0064] In practical applications, in addition to real-time maritime communication translation, recording communication content for subsequent review is crucial for offline large-scale intelligent two-way translation systems aimed at maritime communication support. Therefore, this embodiment of the offline large-scale intelligent two-way translation system for maritime communication support also includes an audio recording unit and a storage unit. The audio recording unit records valid speech upon receiving a recording command from the processing unit 6 and saves the recording to the storage unit. The storage unit stores the recording locally, combining the recording timestamp, first language text, second language text, and the structured format of the second language text translation source. In specific implementation, the audio recording unit communicates with the audio processing unit 1 and the processing unit 6 to record valid speech. The recorded content is stored in the storage unit, which can utilize the PCIe 4.0 channel advantage of the core processor, configuring a 1-2TB NVMe SSD (read / write speed ≥3500MB / s) as the storage medium. The processing unit 6 can be configured to support two recording modes: real-time synchronous recording: after receiving valid voice, it writes to the SSD in WAV format (16kHz / 16-bit) in real time; triggered recording: the user clicks "start recording" to start recording, generates a temporary file when paused, and converts it into a formal file after confirmation and saving.

[0065] Understandably, given the complex environment at sea, communication in emergencies requires less translation time. Therefore, to address sudden emergencies, the processing unit 6 in this application is also used to execute an emergency communication mode upon receiving a mode switching signal. In this emergency communication mode, the translation unit 3 calls a standard language library to translate the first language text into the second language text. The standard language library is a pre-built database storing commonly used phrases for emergency maritime communications. Furthermore, in this mode, the processing unit 6 automatically adjusts the voice generation unit 5 to 90% of the device's maximum volume, adds a 1-second 200Hz beep, and then begins broadcasting the emergency text.

[0066] It is understood that in this application, the first language text is translated into the second language text. The second language text should be in a language that is suitable for local technicians to understand. In order to improve the adaptability of the system, the processing unit 6 is also used to send a language switching instruction to the translation unit 3. The translation unit 3 is used to call the text matching model of the corresponding language according to the language switching instruction after receiving the first language text, and translate the first language text into the second language text.

[0067] The above embodiments provide a detailed description of the offline large-scale intelligent bidirectional translation system for maritime communication support provided in this application. For ease of understanding, this embodiment combines the hardware structure of the assembleable offline large-scale intelligent bidirectional translation system for maritime communication support with the usage method of the offline large-scale intelligent bidirectional translation system for maritime communication support in practical application scenarios to provide a detailed introduction to the technical solution provided in this application. Figure 2 This is a diagram illustrating the user interface of an offline large-scale intelligent two-way translation system for maritime communication support, provided in an embodiment of this application.

[0068] After the offline large-scale intelligent two-way translation system for maritime communication support is powered on, the main program of processing unit 6 starts up. Based on the multi-threading capability of AMD Ryzen7 / Intel i9 processor, it creates four parallel threads for audio acquisition, recording management, speech transcription, and translation processing; it automatically loads the memory of each model; the VAD module of audio processing unit 1 calculates the average energy of the current environmental noise (such as engine noise and sea wind noise) through 3 seconds of silent sampling, and automatically sets the dynamic noise threshold.

[0069] like Figure 2 As shown, users can select the translation timeliness mode ("Quick Translation" or "Accurate Translation") by touching the "Start Recording" button in the GUI, and can click "Start Recording" as needed; the processing unit 6 records the user's instructions and displays status prompts such as "Mode: Accurate Translation | Status: Pending Recording | Recording: Not Started" at the top of the interface.

[0070] The audio processing unit 1 activates the microphone array to acquire an audio stream at 16kHz / 16-bit and transmits it to the memory buffer via the PCIe 4.0 bus.

[0071] If the user has triggered "Start Recording", the audio recording unit and storage unit read audio data from the memory buffer and write it to the NVMeSSD in real time, while generating a temporary metadata file; the VAD module reads audio data from the memory buffer, calculates the energy value per frame (20ms / frame), compares it with the noise threshold, filters out valid speech data blocks, and transmits them to the speech transcription unit 2; the speech transcription unit 2 receives the valid speech data blocks, calls the Faster-Whisper model to perform transcription, and outputs the first language text and confidence score; the processing unit 6 displays the first language text in the corresponding display box in the GUI, and writes the first language text and confidence score to the metadata file of the storage unit, realizing real-time binding of recording and text;

[0072] If "Quick Translation" (Emergency Scenario) is selected: Translation Unit 3 calls the standard language library to directly translate the first language text into the second language text;

[0073] If "Accurate Translation" (normal scenario) is selected: First, the W Ratio algorithm is called to search in the local knowledge base. If the highest similarity is ≥90%, the corresponding translation content is returned as the second language text; if the similarity is <90%, the translation model is called to generate the second language text, and then the BERT error correction model is used to optimize and output the final second language text.

[0074] Processing unit 6 displays the second language text below the corresponding first language text;

[0075] The speech generation unit 5 converts the second language text into a speech signal and broadcasts it through a speaker;

[0076] It is recommended that Unit 4 receive second language text and generate 3-5 suggested replies, which are displayed as blue touch buttons below the GUI text area. Users can trigger the broadcast by clicking any button.

[0077] Emergency communication trigger (optional): When the user clicks the physical "Emergency" button or the preset phrase button in the GUI emergency window (such as "Mayday, Mayday, Mayday", "Air defense alert / air raid alarm"), the processing unit 6 directly transmits the preset emergency text to the natural speech synthesis and playback unit, which executes the broadcast with the highest priority. At the same time, the audio recording unit and storage unit automatically mark the recording as an "emergency event" for easy retrieval later.

[0078] After the user clicks "Stop Recording", the audio recording unit and storage unit convert the temporary recording file into a formal file, and the metadata file is updated to the final version simultaneously;

[0079] The storage unit stores the recorded content locally, combining the recording timestamp, first language text, second language text, and the structured format of the second language text translation source.

[0080] Users can select the target record by checking the checkboxes in the GUI interface and click the "Export" button to choose from two export modes: Text Export: generates a .docx document containing a table of contents and tables; Audio Export: generates a compressed package containing the document and associated audio clips.

[0081] like Figure 2As shown, the GUI interface displays "Maritime Communication System"; it illustrates the current operation interface, including control buttons for three translation processes: "Start Reception," "Pause Reception," and "End Reception"; and a model selection button for "Enter Emergency Model"; the "Recognition and Translation Results" list displays the first language text corresponding to the valid speech and the corresponding second language text; the second language text or reply text can be played through the speech synthesis unit via the "Broadcast" and "Suggest" buttons; the corresponding communication recognition content can be selected through the selection box in front of the text content, and the local knowledge base can be updated and the storage unit can record the information through "Upload Knowledge Base" and "Save as Word Document."

[0082] The foregoing has provided a detailed description of the offline large-scale intelligent two-way translation system for maritime communication support provided by this invention. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this invention.

Claims

1. An offline large-scale intelligent two-way translation system for maritime communication support, characterized in that, include: The audio processing unit (1), the speech transcription unit (2), and the translation unit (3) are included. The audio processing unit (1) is used to collect ambient audio signals and extract valid speech from the ambient audio signals after receiving a start signal; The speech transcription unit (2) converts the effective speech received from the audio processing unit (1) into first language text based on a pre-loaded speech conversion model; The translation unit (3) receives the first language text and calls the text matching model to translate the first language text into the second language text; the text matching model is a model that incorporates the exclusive terminology of Hainan Airlines Communications.

2. The offline large-scale intelligent two-way translation system for maritime communication support according to claim 1, characterized in that, The translation unit (3) receives the first language text and calls the text matching model to translate the first language text into the second language text, including: The translation unit (3) performs text retrieval by calling the local knowledge base based on the first language text; If the similarity between the recorded text in the local knowledge base and the first language text exceeds a preset threshold, the translated text that matches the recorded text will be used as the second language text. If the recorded text is not found, a pre-built translation model is invoked to translate the second language text based on the vocabulary information in the first language text.

3. The offline large-scale intelligent two-way translation system for maritime communication support according to claim 1, characterized in that, Also includes: It is recommended to provide unit (4): The suggestion providing unit (4) is used to call a pre-built conventional suggestion response matching model after receiving the second language text, and generate response text based on the second language text.

4. The offline large-scale intelligent two-way translation system for maritime communication support according to claim 3, characterized in that, Also includes: Speech generation unit (5); The speech generation unit (5) generates speech based on the second language text and the reply text and broadcasts it.

5. The offline large-scale intelligent two-way translation system for maritime communication support according to claim 1, characterized in that, The audio processing unit (1) is used to collect ambient audio signals and extract valid speech from the ambient audio signals after receiving a start signal, including: The audio processing unit (1) collects the ambient audio signal at a preset frequency as the sampling rate and simultaneously calculates the root mean square energy value of each frame of the ambient audio signal. The energy value is compared with the set noise threshold. When the energy value of the environmental audio signal in N consecutive frames is greater than the noise threshold, it is determined to be the valid speech and the data buffer is started. When the energy value of the environmental audio signal in M ​​consecutive frames is not greater than the noise threshold, it is determined to be silent or noise and data buffering is stopped. Both N and M are constants greater than 1.

6. The offline large-scale intelligent two-way translation system for maritime communication support according to claim 2, characterized in that, Also includes: Processing unit (6) and human-computer interaction unit (7) with display function; The processing unit (6) is used to send the start signal to the audio processing unit (1) after receiving the trigger command fed back by the human-computer interaction unit (7). The human-computer interaction unit (7) is also used to receive and display the first language text and the second language text.

7. The offline large-scale intelligent two-way translation system for maritime communication support according to claim 6, characterized in that, The processing unit (6) is also configured to update and save the recorded text that is not found in the local knowledge base to the local knowledge base when a storage signal is received.

8. The offline large-scale intelligent two-way translation system for maritime communication support according to claim 6, characterized in that, It also includes: an audio recording unit and a storage unit; The audio recording unit is used to perform the recording action of the valid voice after receiving the recording instruction sent by the processing unit (6), and save the recording content to the storage unit; The storage unit is used to store the recorded content locally, combining the recording timestamp, the first language text, the second language text, and the structured format of the translation source of the second language text.

9. The offline large-scale intelligent two-way translation system for maritime communication support according to claim 6, characterized in that, The processing unit (6) is also used to execute an emergency communication mode after receiving a mode switching signal; The translation unit (3) in the emergency communication mode is used to call the standard language library and translate the first language text into the second language text; the standard language library is a pre-built database that stores commonly used phrases for emergency communication in maritime communications.

10. The offline large-scale intelligent two-way translation system for maritime communication support according to claim 6, characterized in that, The processing unit (6) is also used to send a language switching instruction to the translation unit (3); The translation unit (3) is used to translate the first language text into the second language text by calling the corresponding language text matching model according to the language switching instruction after receiving the first language text.

Images