In-vehicle voice privacy protection method, system, device and storage medium
By collecting and processing voice signals in the car in the smart cockpit and using disable thesaurus and data desensitization technology, the problem of inaccurate identification and privacy leakage in the on-board environment is solved, and the efficient identification and privacy protection of voice assistants is achieved .
Patent Information
- Application Number
- CN202410433043.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-11
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-04-11
AI Technical Summary
The smart cockpit voice assistant faces problems such as noise interference, severe reverberation, and multi-person voice aliasing in the on-board environment, resulting in inaccurate voice recognition and risk of privacy leakage, especially when discussing sensitive or commercial confidential content in the car, users Worry about data theft.
Using in-vehicle voice privacy protection method, voice signals are collected through microphones for preprocessing, spectrum analysis and feature extraction, voice recognition is recognized using convolutional neural network, and disable thesaurus is set locally for filtering and data desensitization, generating interference signals and Desensitization signals are compounded and uploaded to the cloud only when the user's wishes are met.
Without affecting the quality of voice interaction, the protection of voice privacy in the car is achieved, avoiding the user's sensitive content being monitored and leaked before uploading, and meeting the strict standards of car voice assistants.
Smart Images

Figure CN118197295B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of intelligent cockpit voice privacy protection, and in particular to an in-vehicle voice privacy protection method, system, device and storage medium. Background Art
[0002] With the development of technology, the functions of AI intelligent voice products are becoming more and more abundant, and they are even equipped with voice assistant functions. What's more, the voice assistant can start as soon as the power is turned on. Some of them even continuously monitor the user's voice and store the recognition results in the background library.
[0003] At present, the market share of smart cars is increasing year by year, and voice assistants are becoming more and more common in smart cockpits as a manifestation of intelligence. For example, Xiaopeng's voice assistant has updated the wake-up-free function. After the user powers on the vehicle, the voice assistant will continue to monitor the user's voice and upload it to the cloud for analysis and mapping to execute corresponding actions. For voice commands that are not within the voice assistant's recognition range, the voice assistant will not execute the voice command after recognition.
[0004] In a very complex acoustic environment such as in-car, car voice assistants face the challenges of noise interference, severe reverberation, multi-person voice aliasing, various wind and rain sounds outside the car, and noise interference from other vehicles. This is a difficult problem for accurate voice recognition, voice quality enhancement, and high-quality voice interaction. At the same time, in-car voice assistants may involve privacy issues, which is also a challenge currently faced by consumers. Sometimes some of the topics we discuss in the car involve privacy or business secrets, and we hope to talk in a private space and keep the content of the conversation completely confidential; but in a smart cockpit with a voice assistant function, the voice of the current speaker will be monitored and even uploaded to the cloud. As a voice assistant, as long as the voice is recognized, it will be uploaded to the parser for a series of analysis and calculations. As long as it is uploaded to the cloud, there will be confusion about data security, and there will be concerns about data leakage or theft of business secrets.
[0005] Car voice assistants are different from other devices. To meet the stringent automotive standards, the model size and real-time rate of the in-vehicle voice system need to be controlled within a relatively low range, with low CPU usage and fast overall response time. The requirements for recognition effect, broadcast experience, response time, resource consumption and other indicators are higher than those of ordinary recognition models.
[0006] In view of this, we need a method that can meet the stringent standards of automotive voice assistants while addressing the issue of in-car voice privacy protection. Summary of the invention
[0007] In view of the strict design standards of car voice assistants and the need to solve the problem of privacy leakage, a method, system, device and storage medium for in-vehicle voice privacy protection are proposed.
[0008] In order to achieve the above object, the present invention is implemented by the following technical solutions:
[0009] In one aspect, the present invention provides a method for protecting voice privacy in a vehicle, the method comprising:
[0010] Collect voice signals from the sound source in the car through a microphone;
[0011] Preprocess the collected speech signals, including noise removal, pre-emphasis, and frame division;
[0012] Perform spectrum analysis on the preprocessed speech signal and extract features from the speech spectrum graph through a convolutional neural network;
[0013] The decoder starts decoding, inputs the extracted features into the speech recognition model, and parses candidate words through the acoustic model and dictionary;
[0014] The candidate word is matched with the local banned word library to determine whether the candidate word is a banned word. If so, it is banned, filtered and data desensitized without any interface feedback. At the same time, an interference voice signal is generated, which is combined with the desensitized voice signal to form a desensitized voice interference signal and transmitted to the language model; if not, the result that meets the user's expression content is directly obtained through the language model and stored in the cloud;
[0015] Output recognition results, that is, the results in the form of converted text or instructions.
[0016] As a preferred solution of the present invention, the noise removal specifically comprises: extracting the frequency, time domain, and energy characteristics of the sound in the speech signal, distinguishing the human voice and other sounds in the speech signal by comparative analysis of the characteristics, and removing the noise in the speech signal by filtering and noise reduction algorithms, wherein the noise reduction algorithms include Fourier transform based on the frequency domain, wavelet transform, and signal smoothing method based on the time domain;
[0017] The pre-emphasis specifically refers to strengthening the energy of high-frequency components and reducing the energy of low-frequency components, so that speech signals of different frequencies can be equalized during the signal processing process.
[0018] As a preferred solution of the present invention, the feature extraction specifically includes: extracting features used to describe the speech signal from the speech signal through linear prediction cepstral coefficients and Mel cepstral coefficients methods, and the features include: frame energy, pitch, fundamental frequency, tone period, resonance peaks, harmonic structure and vocal tract characteristics.
[0019] As a preferred solution of the present invention, the local banned word library is in the form of a form, stored in a local database, and supports users to create, modify, save and delete banned word library contents.
[0020] As a preferred solution of the present invention, the method for generating a speech interference signal specifically includes: performing time reversal and random short-time phase spectrum operations on each frame of the framed speech signal, overlapping and adding each frame signal, and combining irrelevant interference signals in a speech interference library to generate the speech interference signal; the irrelevant interference signal includes bird calls, piano sounds, drum beats or flowing water sounds.
[0021] On the other hand, the present invention also provides an in-vehicle voice privacy protection system, the system comprising:
[0022] The sound source acquisition module is used to collect the voice signal of the sound source in the car and send it to the pre-processing module;
[0023] The preprocessing module is used to preprocess the collected speech signals, including noise removal, pre-emphasis, and frame division;
[0024] A feature extraction module is used to perform spectrum analysis on the preprocessed speech signal and extract features from the speech spectrum through a convolutional neural network; the features include: frame energy, pitch, fundamental frequency, tone period, formant, harmonic structure and vocal tract characteristics;
[0025] A decoder module, used to generate speech recognition results, including a candidate word parsing unit, a banned word matching unit, a data desensitization unit, an interference signal unit, and a language model output unit;
[0026] A recognition result output module is used to output the converted result in the form of text or instructions;
[0027] The local banned word library is used to store the banned words preset by the user in the local database. The local banned word library is in the form of a form, which supports the user to create, modify, save and delete the banned word library content.
[0028] As a preferred solution of the present invention, the candidate word parsing unit is used to input the extracted features into the speech recognition model, and parse the candidate words through the acoustic model and the dictionary;
[0029] The banned word matching unit is used to match the candidate word with a local banned word library to determine whether the candidate word is a banned word;
[0030] The data desensitization unit is used to perform a filtering and data desensitization process on the banned words to obtain a desensitized voice signal;
[0031] The interference signal unit is used to generate a speech interference signal, perform time reversal and random short-time phase spectrum operations on each frame of the framed speech signal, overlap and add each frame signal, and generate the speech interference signal by combining the irrelevant interference signal of the speech interference library; the irrelevant interference signal includes bird calls, piano sounds, drum beats or running water sounds;
[0032] The signal combining unit is used to combine the interference voice signal and the desensitized voice signal into a desensitized voice interference signal and transmit it to the language model;
[0033] The speech model output unit is used to directly obtain a result that matches the user's expression content through a language model and store it in the cloud.
[0034] The present invention also provides a device, comprising: one or more processors; a memory for storing one or more programs; and when one or more of the programs are executed by one or more of the processors, an in-vehicle voice privacy protection method as described above is implemented.
[0035] The present invention also provides a storage medium containing computer executable instructions, which are used to execute the in-vehicle voice privacy protection method as described above when executed by a computer processor.
[0036] Compared with the prior art, the beneficial effects of the present invention are: without affecting the structure of the existing voice interaction system, a banned word strategy is introduced; when the voice assistant is running in the background, the MIC in the car starts to collect the user's voice, and a local banned word library, namely a banned word funnel, is added when uploading it to the cloud for analysis, and filtering and data desensitization are performed when identifying the complete semantics, and interference signals are added at the same time, and the parsed content is saved; the content that the user does not want to be monitored during this time period can be filtered and analyzed as soon as possible before uploading it to the cloud, thereby avoiding the leakage of user privacy. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:
[0038] Figure 1 This is the basic framework diagram of the existing intelligent cockpit voice interaction system;
[0039] Figure 2 is a flow chart of the method of the present invention;
[0040] Figure 3 It is a system structure diagram of the present invention. DETAILED DESCRIPTION
[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.
[0042] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0043] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.
[0044] In the smart cockpit scenario, voice interaction has become a very critical human-computer interaction method. From the user's perspective, the core value of voice interaction is mainly reflected in freeing the user's hands, making the interaction between people and machines more efficient and convenient. However, the process from the user issuing a voice command to realizing the interaction with the smart cockpit is not as simple as its noun describes. In order to ensure the quality of voice interaction, existing car voice assistants start when powered on, and some even continuously monitor the user's voice and store the recognition results in the background library. However, this may involve privacy protection issues, which is a challenge currently faced by consumers. How to solve the problem of voice privacy protection in the car without affecting the quality of voice interaction is a problem that the present invention needs to solve.
[0045] like Figure 1 As shown in the figure, the voice interaction system of the smart cockpit is a highly complex integrated system, which not only requires the coordination of precise hardware and software, but also requires professional operation management to ensure its reliability and stability. In general, the entire system can be divided into four components: hardware layer, service layer, application layer and operation management platform.
[0046] Hardware layer: In the voice interaction system, the hardware layer is the physical basis of the smart cockpit voice interaction system. The key hardware devices include input / output devices and audio processing chips. The input / output devices are responsible for collecting the user's voice commands and feedback information, mainly including: microphone arrays, speakers, cameras, sensors, lights, etc. The chip part is mainly responsible for the processing and analysis of audio signals, mainly including digital signal processors (DSP), audio decoders, audio amplifiers, etc.
[0047] Service layer: The service layer is the core of the intelligent cockpit voice interaction system. It is responsible for processing and analyzing voice, image and other information, and providing the necessary feedback and response. It mainly includes modules such as natural language processing (NLP) engine, speech recognition engine, speech synthesis engine, voiceprint recognition, cloud service, API service, business logic processing service, etc.
[0048] Application layer: The application layer refers to specific applications provided to users based on the core capabilities provided by the service layer combined with the actual needs of users, to help users achieve specific functional control through voice interaction. For example, applications such as air conditioning control, seat control, and window control in the body control module, and entertainment, communication, and navigation applications in the infotainment module.
[0049] Operation management platform: When users use the voice interaction system, they will generate a large amount of user behavior data and audio, text, and image data. The operation management platform provides data support for the continuous optimization of the voice interaction system and various AI models through statistical analysis of these data. From a business perspective, the operation management platform is mainly divided into two core modules: user data statistical analysis and model data operation.
[0050] The user data statistics and analysis module can collect statistics and analyze the behavior data of real car users using the voice interaction system, thereby generating analysis reports of different dimensions and granularities. These reports can help us gain a deeper understanding of user usage habits and preferences, promptly discover and solve system problems, and provide data basis for the optimization of the voice interaction system.
[0051] The model data operation module can produce the data required by each AI model by regularly recycling, collecting and annotating the large amount of audio, text and image data generated by users when using the voice interaction system, thereby providing data support for model training.
[0052] Under the premise of not changing the structure of the above-mentioned voice interaction system, the present invention uses a banned word strategy to perform filtering and data desensitization when recognizing the complete semantics, thereby preventing the user's private data from being leaked in the first place.
[0053] like Figure 2 As shown in FIG. 1 , an embodiment of the present invention is provided, which provides a method for protecting voice privacy in a vehicle, and specifically includes the following steps:
[0054] S1: collects the voice signal of the sound source in the car through the microphone;
[0055] S2: pre-processing the collected speech signal, including noise removal, pre-emphasis, frame division, windowing, speech signal energy normalization, frequency filtering, dynamic features, etc.;
[0056] Due to the influence of user environment factors, the sound recorded by the MIC device may include various noises in addition to the human voice. Therefore, in order to ensure the accuracy of speech recognition, the noise part of the original speech signal needs to be processed before recognition.
[0057] In a specific embodiment, noise removal may include: extracting frequency, time domain, energy and other characteristics of sound in the speech signal, distinguishing human voice from other sounds in the speech signal through comparative analysis of the characteristics, and removing noise in the speech signal through filtering and noise reduction algorithms. The noise reduction algorithms may include Fourier transform based on frequency domain, wavelet transform and signal smoothing method based on time domain.
[0058] During the voice input process, due to factors such as the environment and distance, the MIC recorded sound may experience distortion such as high-frequency attenuation and low-frequency gain, which will affect the subsequent voice recognition results.
[0059] For example, if the user's voice content is "apple", due to the attenuation of high-frequency signals, only the strong signals of "p" and "l" may be left in the recorded sound, which will cause the speech recognition system to mistakenly think that "pl" is said instead of "apple". To address this phenomenon, pre-emphasis can enhance the energy of high-frequency components and reduce the energy of low-frequency components, so that audio signals of different frequencies can be equalized during the signal processing process, thereby improving the accuracy of speech recognition.
[0060] The original speech signal is a continuous waveform, a signal that changes rapidly in both time and frequency. In the process of speech recognition, if the continuous and long speech is directly processed, the difficulty of calculation will increase and the accuracy of recognition will be reduced. Therefore, in order to improve the accuracy of speech recognition results, the continuous and long speech signal needs to be divided into several frames of fixed length. After the framing, the spectrum of the signal in each frame will change slowly and stably.
[0061] Taking "turn on the air conditioner" as an example, assuming that the recorded voice duration is 2 seconds and the sampling rate is 16000 Hz, then the original voice signal is a one-dimensional vector with a length of 32000. If speech recognition is performed directly on this sound information, the amount of calculation will be very large, and because the frequency and amplitude of the voice signal change very quickly, it is difficult to perform effective feature extraction.
[0062] S3: Perform spectrum analysis on the preprocessed speech signal and extract features from the speech spectrum graph through a convolutional neural network;
[0063] The speech signal after front-end processing such as denoising, pre-emphasis, and framing cannot be used directly for recognition. It needs to be transformed into the frequency domain and then extracted from the speech signal using methods such as Linear Prediction Cepstral Coefficients (LPCC) and Mel Cepstral Coefficients (MFCC) to describe the various features of the speech signal so that the speech recognition model can better analyze and distinguish it. These features include: frame energy, pitch (tone, tone, etc.), fundamental frequency, tone period, resonance peaks, harmonic structure, vocal tract characteristics, etc.
[0064] S4: The decoder starts decoding, inputs the extracted features into the speech recognition model, and parses candidate words through the acoustic model and dictionary;
[0065] The acoustic model is responsible for extracting and processing the features of the speech signal, generating a series of feature vectors, and then using these feature vectors to calculate the score of each possible phoneme, and selecting the most likely phoneme sequence based on the score; the Chinese characters (words) or words corresponding to the phonemes are identified through the dictionary. The dictionary in the speech recognition system includes a series of words and their corresponding phoneme sequences. These phoneme sequences reflect the phonetic features and pronunciation of the words in the speech signal. By matching the actual pronunciation of the speech signal with the pronunciation stored in the dictionary, the speech recognition system can infer the words spoken by the speaker.
[0066] Create a local banned word library on the vehicle computer. The local banned word library is in the form of a form and stored in the local database. It supports users to create, modify, save and delete the banned word library content;
[0067] When the decoder completes its work and parses the candidate words, it calls the local banned word library for matching, determines whether the candidate word is a banned word, and if so, disables it, performs filtering and data desensitization processing (mask shielding: use * to mask part of the data, such as retaining the first 6 digits of the ID card representing the regional information, and replace the rest with *. The masked part can be adjusted as needed), without any interface feedback. At the same time, the framed voice signal is time-reversed and randomized for each frame. Each frame signal is overlapped and added, and the irrelevant interference signal of the voice interference library is combined to generate a voice interference signal, which is combined with the desensitized voice signal to form a desensitized voice interference signal and transmitted to the language model; irrelevant interference signals can include bird calls, piano sounds, drum beats or running water sounds, etc.
[0068] If it is not a banned word, the language model is used to directly obtain the result that matches the user's expression and store it in the cloud;
[0069] The role of the speech model is to predict the probability of a word or sentence appearing by counting the relationship and probability between words in the text, so as to score, sort and filter the multiple identified text candidate results. Ultimately, the result with the highest score is the result that the system believes best matches the user's expression.
[0070] S5: Output the recognition result, that is, the result in the converted text or instruction form.
[0071] like Figure 3 As shown, another embodiment of the present invention provides an in-vehicle voice privacy protection system, which is mainly divided into the following five modules:
[0072] The sound source acquisition module is used to collect the voice signal of the sound source in the car and send it to the pre-processing module;
[0073] The preprocessing module is used to preprocess the collected speech signals, including noise removal, pre-emphasis, and frame division;
[0074] The feature extraction module is used to perform spectrum analysis on the preprocessed speech signal and extract features from the speech spectrum through a convolutional neural network; the features include: frame energy, pitch, fundamental frequency, tone period, formant, harmonic structure, and vocal tract characteristics;
[0075] A decoder module, used to generate speech recognition results, including a candidate word parsing unit, a banned word matching unit, a data desensitization unit, an interference signal unit, and a language model output unit;
[0076] The candidate word parsing unit is used to input the extracted features into the speech recognition model, and parse the candidate words through the acoustic model and the dictionary; the banned word matching unit is used to match the candidate words with the local banned word library to determine whether the candidate words are banned words; the data desensitization unit is used to filter the banned words and perform data desensitization processing to obtain a desensitized speech signal; the interference signal unit is used to generate a speech interference signal, perform time reversal and random short-time phase spectrum operations on each frame of the framed speech signal, overlap and add each frame signal, and generate a speech interference signal in combination with irrelevant interference signals in the speech interference library; irrelevant interference signals include bird calls, piano sounds, drum beats or running water sounds; the signal compounding unit is used to compound the interference speech signal with the desensitized speech signal into a desensitized speech interference signal and transmit it to the language model; the speech model output unit is used to directly obtain the result that conforms to the user's expression content through the language model, and store it in the cloud;
[0077] A recognition result output module is used to output the converted result in the form of text or instructions;
[0078] The system also includes a local banned word library for storing user preset banned words in a local database. The local banned word library is in form of a form with a corresponding interactive UI interface, which supports users to create, modify, save and delete banned word library contents.
[0079] It should be understood that the various parts of the present application can be implemented with hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented with software or firmware stored in a memory and executed by a suitable instruction execution system. All or part of the steps of the above embodiment method can be completed by instructing the relevant hardware through a program, which can be stored in a computer-readable storage medium, and when the program is executed, it includes one of the steps of the method embodiment or a combination thereof.
[0080] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the above-mentioned integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium. The storage medium may be a read-only memory, a disk or an optical disk, etc., for executing the above-mentioned in-vehicle voice privacy protection method.
[0081] To sum up, the present invention introduces a banned word strategy without affecting the structure of the existing voice interaction system. When the voice assistant is running in the background and the MIC in the car collects the user's voice, a local banned word library, namely a banned word funnel, is added when uploading the cloud for analysis. When the complete semantics is recognized, filtering and data desensitization are performed, and interference signals are added at the same time, and the parsed content is saved. The content that the user does not want to be monitored during this time period can be filtered and analyzed as soon as possible before uploading to the cloud, thereby avoiding the leakage of user privacy.
[0082] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A method for protecting voice privacy in a vehicle, characterized in that: The method comprises: Collect voice signals from the sound source in the car through a microphone; Preprocess the collected speech signals, including noise removal, pre-emphasis, and frame division; The noise removal is specifically: extracting the frequency, time domain, and energy characteristics of the sound in the speech signal, distinguishing the human voice and other sounds in the speech signal by comparative analysis of the characteristics, and removing the noise in the speech signal by filtering and noise reduction algorithms, wherein the noise reduction algorithms include Fourier transform based on frequency domain, wavelet transform, and signal smoothing method based on time domain; The pre-emphasis specifically refers to strengthening the energy of high-frequency components and reducing the energy of low-frequency components, so that speech signals of different frequencies can be equalized during the signal processing process; Perform spectrum analysis on the preprocessed speech signal and extract features from the speech spectrum graph through a convolutional neural network; The feature extraction specifically includes: extracting features used to describe the speech signal from the speech signal by using the linear prediction cepstral coefficient and Mel cepstral coefficient method, the features including: frame energy, pitch, fundamental frequency, tone period, formant, harmonic structure and vocal tract characteristics; The decoder starts decoding, inputs the extracted features into the speech recognition model, and parses candidate words through the acoustic model and dictionary; The candidate word is matched with the local banned word library to determine whether the candidate word is a banned word. If so, it is banned, filtered and data desensitized without any interface feedback. At the same time, a voice interference signal is generated, which is combined with the desensitized voice signal to form a desensitized voice interference signal and transmitted to the language model; if not, the result that meets the user's expression content is directly obtained through the language model and stored in the cloud; The local banned word library is in the form of a form, stored in a local database, and supports users to create, modify, save and delete banned word library contents; The method for generating a speech interference signal specifically comprises: performing time reversal and random short-time phase spectrum operations on each frame of the framed speech signal, overlapping and adding each frame signal, and combining an irrelevant interference signal of a speech interference library to generate the speech interference signal; the irrelevant interference signal comprises bird calls, piano sounds, drum beats or running water sounds; Output recognition results, that is, the results in the form of converted text or instructions.
2. An in-vehicle voice privacy protection system, based on an in-vehicle voice privacy protection method as claimed in claim 1, characterized in that: The system comprises: The sound source acquisition module is used to collect the voice signal of the sound source in the car and send it to the pre-processing module; The preprocessing module is used to preprocess the collected speech signals, including noise removal, pre-emphasis, and frame division; A feature extraction module is used to perform spectrum analysis on the preprocessed speech signal and extract features from the speech spectrum through a convolutional neural network; the features include: frame energy, pitch, fundamental frequency, tone period, formant, harmonic structure and vocal tract characteristics; A decoder module, used to generate speech recognition results, including a candidate word parsing unit, a banned word matching unit, a data desensitization unit, an interference signal unit, and a language model output unit; The candidate word parsing unit is used to input the extracted features into the speech recognition model and parse the candidate words through the acoustic model and the dictionary; The banned word matching unit is used to match the candidate word with a local banned word library to determine whether the candidate word is a banned word; The data desensitization unit is used to perform a filtering and data desensitization process on the banned words to obtain a desensitized voice signal; The interference signal unit is used to generate a speech interference signal, perform time reversal and random short-time phase spectrum operations on each frame of the framed speech signal, overlap and add each frame signal, and generate the speech interference signal by combining the irrelevant interference signal of the speech interference library; the irrelevant interference signal includes bird calls, piano sounds, drum beats or running water sounds; The signal combining unit is used to combine the voice interference signal and the desensitized voice signal into a desensitized voice interference signal and transmit it to the language model; The language model output unit is used to directly obtain a result that matches the user's expression through the language model and store it in the cloud; A recognition result output module is used to output the converted result in the form of text or instructions; The local banned word library is used to store the banned words preset by the user in the local database. The local banned word library is in the form of a form, which supports the user to create, modify, save and delete the banned word library content.
3. A device, characterized in that: The device includes: one or more processors; a memory for storing one or more programs; when one or more of the programs are executed by one or more of the processors, an in-vehicle voice privacy protection method as claimed in claim 1 is implemented.
4. A storage medium containing computer executable instructions, characterized in that: The computer executable instructions are used to execute the in-vehicle voice privacy protection method as claimed in claim 1 when executed by a computer processor.
Citation Information
Patent Citations
Automatic speech recognition method and system based on artificial intelligence
CN112951237A
Voice data processing method, electronic equipment and computer readable storage medium
CN113033191A
Smart home interaction method and device based on Internet of Things
CN114743554A
Privacy protection method for artificial intelligence voice partner training
CN117648717A
Privacy control in a connected environment based on speech characteristics
US20180213396A1
