Voice imitation system based on AI

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an AI-based voice simulation system, which comprises a voice acquisition module, a voice processing module, a voice synthesis module and a voice feedback module, and is characterized in that firstly, the voice acquisition module is used for acquiring real-time and historical audio data of a user; secondly, the voice processing module is used for converting the data into a multi-dimensional vector and constructing an initial generative adversarial network model; then, a speech synthesis module training model generates imitated speech, and the imitated speech is compared with original sound of a user for adjustment; and finally, the voice feedback module further optimizes the multi-dimensional vector according to user feedback, and retrains the initial generative adversarial network model until a satisfactory imitation effect is achieved. According to the invention, through deep learning and user feedback loop optimization, highly personalized voice simulation is realized, and user experience and satisfaction are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-based voice imitation system. Background Art

[0002] With the development of integrated circuit technology, more complex electronic sound generators have emerged. For example, the series / parallel hybrid formant synthesizer released by KLATT in 1980 significantly improved the naturalness of speech synthesis.

[0003] Since the beginning of the 21st century, with the development of deep learning technology, speech synthesis technology has made significant progress. Modern speech synthesis systems can produce more natural and fluent speech and are widely used in fields such as virtual assistants, barrier-free communication, language learning, and media entertainment.

[0004] Patent document with publication number CN106571145A discloses a voice imitation method and device, which includes: obtaining a training voice of a source user and a plurality of training voices of imitating users respectively; determining a conversion rule for converting the training voice of the source user into the training voice of the target user based on the training voice of the source user and the training voice of a target user among the plurality of imitating users; collecting voice information of the source user; and converting the voice information of the source user into the imitating voice of the target user based on the determined conversion rule for converting the training voice of the source user into the training voice of the target user.

[0005] This shows that the following problem exists: because the conversion rules in the prior art oversimplify the complexity of voice imitation, the generated voice lacks personalized timbre details. Summary of the Invention

[0006] The purpose of the present invention is to provide an AI-based voice imitation system to solve the problem in the prior art that conversion rules oversimplify the complexity of voice imitation, resulting in the generated voice lacking personalized timbre details.

[0007] To achieve the above objectives, the present invention provides an AI-based voice imitation system, comprising:

[0008] The voice acquisition module is used to obtain the user's audio signal and voice information, and also to obtain the user's historical voice information and historical audio signal;

[0009] a speech processing module, connected to the speech acquisition module, for combining the speech information, the audio signal, the historical speech information, and the historical audio signal to obtain a multidimensional vector, and constructing an initial generative adversarial network model with the multidimensional vector and a predefined language database;

[0010] a speech synthesis module, connected to the speech processing module, configured to train the initial generative adversarial network model to obtain an initial simulated speech, compare the initial simulated speech with the speech information to obtain an actual similarity score, and compare the actual similarity score with a preset standard similarity score to adjust the multidimensional vector to obtain a first adjustment result;

[0011] A speech feedback module is connected to the speech synthesis module and is used to obtain a feedback result based on the user's feedback on the initial imitation speech, adjust the multidimensional vector based on the feedback result to obtain a second adjustment result, merge the first adjustment result and the second adjustment result to obtain a target multidimensional vector, and retrain the initial generative adversarial model based on the target multidimensional vector to obtain a target imitation speech.

[0012] Furthermore, the voice processing module includes:

[0013] a preprocessing unit, configured to perform noise elimination and echo suppression on the voice information, the audio signal, the historical voice information, and the historical audio signal to obtain a first processing result, and to segment the audio signal and the historical audio signal into a plurality of small segments to obtain a second processing result;

[0014] a feature extraction unit connected to the preprocessing unit, configured to capture time sequence information of the first processing result to extract features of intonation and rhythm to obtain a first feature result, and further configured to extract features of timbre and pronunciation from the second processing result to obtain a second feature result;

[0015] a parameter generating unit, connected to the feature extracting unit, and configured to combine the first feature result and the second feature result to obtain the multidimensional vector;

[0016] A model construction unit is connected to the parameter generation unit and is used to construct the initial generative adversarial network model using the multidimensional vector and a predefined language database as input parameters.

[0017] Furthermore, the parameter generation unit includes:

[0018] a feature normalization subunit, configured to perform normalization processing on the first feature result and the second feature result to obtain a third processing result and a fourth processing result respectively;

[0019] A feature combination subunit is connected to the feature normalization subunit and is used to fuse the third processing result and the fourth processing result to obtain the multi-dimensional vector.

[0020] Furthermore, the model building unit includes:

[0021] A database interface subunit, used to connect to a predefined voice database to obtain voice data;

[0022] A model construction subunit is connected to the database interface subunit and is used to perform pre-training based on the voice data and the multidimensional vector to construct the initial generative adversarial network model.

[0023] Furthermore, the speech synthesis module includes:

[0024] A model training unit, configured to train the initial generative adversarial model to obtain the initial imitation speech;

[0025] a similarity calculation unit, connected to the model training unit, for comparing the initial simulated voice with the voice information to calculate the actual similarity score;

[0026] a score comparison unit, connected to the similarity comparison unit, for comparing the actual similarity score with a preset standard similarity score to obtain a comparison result;

[0027] A vector adjustment unit is connected to the score comparison unit and is used to adjust the multi-dimensional vector according to the comparison result to obtain the first adjustment result.

[0028] Furthermore, the similarity calculation unit includes:

[0029] a time alignment subunit, configured to align the initial simulated speech and the speech information in time according to dynamic time warping to obtain an alignment result;

[0030] a feature matching subunit, connected to the time alignment subunit, for performing feature matching on the alignment result to obtain a matching result;

[0031] a distance calculation subunit, connected to the matching result subunit, for calculating the Euclidean distance between the matching results to obtain a calculation result;

[0032] a similarity accumulation subunit, connected to the distance calculation subunit, for accumulating the calculation results to obtain an accumulation result;

[0033] The score normalization subunit is connected to the similarity accumulation subunit and is used to normalize the accumulation result to obtain the actual similarity score.

[0034] Furthermore, the scoring comparison unit includes:

[0035] a score alignment subunit, configured to determine the alignment of the actual similarity score and the standard similarity score in terms of dimension and format to obtain an alignment result;

[0036] The score calculation subunit is connected to the score alignment subunit and is used to calculate the difference between the actual similarity score and the standard similarity score according to the alignment result to obtain the comparison result.

[0037] Furthermore, the voice feedback module includes:

[0038] A feedback analysis unit, configured to perform text or voice analysis on the feedback result to obtain an analysis result;

[0039] a vector adjustment unit, connected to the feedback analysis unit, configured to map the analysis result to a multi-dimensional vector space to determine an adjusted vector dimension to obtain a second adjustment result;

[0040] a result integration unit connected to the vector adjustment unit, configured to combine the first adjustment result and the second adjustment result to form a target multi-dimensional vector;

[0041] A model retraining unit is connected to the result integration unit and is used to prepare training data according to the target multidimensional vector and retrain the initial generative adversarial network model according to the prepared data to obtain the target imitation speech.

[0042] Furthermore, the vector adjustment unit includes:

[0043] A dimension mapping subunit, used to adjust the vector dimension according to the feedback result to obtain a mapping result;

[0044] an adjustment strategy subunit, connected to the dimension mapping subunit, for adjusting the weight of the corresponding vector dimension according to the mapping result to obtain a third adjustment result;

[0045] The vector calculation subunit is connected to the adjustment strategy subunit and is used to calculate the third adjustment result to obtain the second adjustment result.

[0046] Furthermore, the model retraining unit includes:

[0047] A model configuration subunit, configured to adjust the parameters of the initial generative adversarial network model according to new training data and objectives to obtain a configuration result;

[0048] The training execution subunit is connected to the model configuration subunit and is used for performing training according to the configuration result to obtain the target imitation speech.

[0049] Compared with the prior art, the beneficial effect of the present invention is that the present invention combines the user's real-time and historical voice data through the voice acquisition module to generate highly personalized imitation voice, meeting the user's personalized needs for voice imitation. The voice processing module converts complex voice information into multi-dimensional vectors, which is convenient for model processing and analysis. The initial GAN model is constructed using multi-dimensional vectors and a predefined language database, providing an efficient starting point for voice synthesis. The voice synthesis module achieves accurate imitation of the user's voice through multi-dimensional vector combination and generative adversarial network model training, and the actual similarity score is close to the preset standard, which improves the accuracy of voice imitation. The voice feedback module introduces a user feedback mechanism, and iteratively optimizes according to the user's feedback on the initial imitation voice to ensure that the imitation voice continues to approach user expectations and achieve continuous improvement. The system can be applied to a variety of scenarios such as voice assistants, virtual characters, and voice conversion, improving the user experience and interactivity of these applications.

[0050] In particular, the preprocessing unit improves the quality of the speech signal through noise cancellation and echo suppression, reducing noise and echo interference. The feature extraction unit extracts timbre and pronunciation features, providing a basis for generating imitation speech with the target user's timbre and pronunciation characteristics. This helps achieve highly personalized speech imitation and enhances the user experience. The parameter generation unit combines intonation, rhythm, timbre, and pronunciation features into a multidimensional vector to form a complete speech feature representation. This provides rich input information for building a generative adversarial network model, facilitating the generation of high-quality imitation speech. The model construction unit utilizes the multidimensional vector and a predefined language database to construct an initial GAN model, providing a basic framework for speech generation.

[0051] In particular, the feature normalization unit ensures the consistency of feature vectors through normalization, facilitating stable model training. The feature combination subunit fuses multiple normalized feature results into a multidimensional vector, preserving important information from the original speech data, such as intonation, rhythm, timbre, and pronunciation, providing a comprehensive information foundation for generating high-quality imitation speech.

[0052] In particular, the database interface subunit allows the model to easily access predefined voice databases, eliminating the need for manual data import and improving data usability. The model building subunit, through multi-dimensional vector input, enables the model to learn personalized voice features and generate imitation voices that are more suitable for the target user.

[0053] In particular, the model training process enhances the model's generative capabilities, enabling it to better mimic the voice characteristics of a specific user. The similarity calculation unit provides a quantitative similarity score that intuitively reflects the degree of similarity between the imitated voice and the original user's voice.

[0054] In particular, the time alignment subunit, using the dynamic time warping (DTW) algorithm, effectively aligns speech signals of varying lengths, ensuring the accuracy of subsequent feature matching. The feature matching subunit distinguishes the voice characteristics of different users, enabling more personalized voice imitation. This effective feature matching algorithm reduces unnecessary computation and improves the overall system processing efficiency. The distance calculation subunit converts feature matching results into Euclidean distance, providing an intuitive and quantitative representation of similarity. The similarity accumulation subunit, by accumulating matching differences across frames, comprehensively assesses the similarity of the entire speech segment, preventing poor local matching from impacting the overall performance. The score normalization subunit normalizes the accumulated results to a score range of 0 to 1, ensuring a uniform and intuitively interpretable similarity score. This allows users to more easily understand the similarity score and better evaluate the effectiveness of the voice imitation.

[0055] In particular, the score alignment subunit ensures comparability between the actual and standard similarity scores by aligning their dimensions and formats, enabling effective comparative analysis. The score calculation subunit quantifies the score comparison results by calculating the difference between the actual and standard similarity scores, facilitating comparison and evaluation.

[0056] In particular, the feedback analysis unit extracts key information from user feedback and identifies user intent through text or speech analysis. The vector adjustment unit determines the direction of model adjustment by analyzing the vector dimensions corresponding to user feedback. The model retraining unit continuously collects user feedback and retrains the model, enabling continuous improvement of the model and enhancing the quality of speech generation.

[0057] In particular, the Dimension Mapping subunit accurately identifies and locates the vector dimensions that require adjustment based on feedback, ensuring targeted adjustments. The Adjustment Strategy subunit effectively optimizes speech generation quality by adjusting the weights of vector dimensions, ensuring that the generated speech better meets user needs. The Vector Calculation subunit normalizes the results of the third processing step to ensure that the adjusted vectors maintain consistent format with other vectors, facilitating subsequent processing.

[0058] In particular, the model configuration subunit can personalize the parameters of the initial generative adversarial network model based on new training data and user feedback targets, making the model more tailored to the needs of specific users. The model retraining unit significantly improves the quality of generated speech through the collaborative work of configuration and training, making it more natural and more in line with user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic diagram of the structure of an AI-based voice imitation system provided in an embodiment of the present invention;

[0060] Figure 2 A schematic diagram of the structure of a voice processing module in an AI-based voice imitation system provided in an embodiment of the present invention;

[0061] Figure 3 A schematic diagram of the structure of a speech synthesis module in an AI-based speech imitation system provided in an embodiment of the present invention;

[0062] Figure 4 A schematic diagram of the structure of a voice feedback module in an AI-based voice imitation system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0064] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0065] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside", and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0066] Furthermore, it should be noted that, in the description of the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0067] See also Figure 1 As shown, an embodiment of the present invention provides an AI-based voice imitation system, including:

[0068] The voice collection module 10 is used to obtain the user's audio signal and voice information, and also to obtain the user's historical voice information and historical audio signal;

[0069] a speech processing module 20 connected to the speech acquisition module 10, configured to combine the speech information, the audio signal, the historical speech information, and the historical audio signal to obtain a multidimensional vector, and construct an initial generative adversarial network model using the multidimensional vector and a predefined language database;

[0070] a speech synthesis module 30 connected to the speech processing module 20, configured to train the initial generative adversarial network model to obtain an initial simulated speech, compare the initial simulated speech with the speech information to obtain an actual similarity score, and compare the actual similarity score with a preset standard similarity score to adjust the multidimensional vector to obtain a first adjustment result;

[0071] The speech feedback module 40 is connected to the speech synthesis module 30 and is used to obtain a feedback result based on the user's feedback on the initial imitation speech, adjust the multidimensional vector based on the feedback result to obtain a second adjustment result, merge the first adjustment result and the second adjustment result to obtain a target multidimensional vector, and retrain the initial generative adversarial model based on the target multidimensional vector to obtain a target imitation speech.

[0072] Specifically, throughout the entire voice imitation process, all user-recorded voice data utilizes advanced encryption algorithms during transmission and storage to prevent unauthorized third-party access. The system implements strict data access control, ensuring that data can only be accessed through user-authorized operations. To minimize the risk of data leakage, initial data processing and feature extraction are performed locally on the user's device whenever possible. Data is only transmitted to a secure server for further processing when necessary and with explicit user authorization. During data upload and storage, the system anonymizes user identity information to ensure that any data analysis results cannot be traced back to an individual. A data security monitoring module is also implemented to monitor data access and transmission in real time, providing prompt alerts upon detection of anomalies. All data is regularly backed up to prevent data loss due to system failures or unexpected events.

[0073] Specifically, a predefined language database containing various voice features, pitch, speaking rate, and other information is constructed to provide basic data for subsequent voice processing and imitation. The voice acquisition module collects the user's audio signals in real time using a microphone array. Voice recognition technology is used to extract voice information, such as pitch, timbre, and speaking rate, from the real-time audio. The user's historical voice information and audio signals are retrieved from the database. Feature extraction, including Mel-Frequency Cepstral Coefficients (MFCCs), pitch, and energy, is performed on the real-time and historical audio signals. Feature extraction, such as pitch variation and speaking rate variation, is performed on the real-time and historical voice information. The extracted real-time and historical features are combined into a multidimensional vector to form a vector that comprehensively represents the user's voice characteristics. Using the multidimensional vector and the predefined language database, an initial generative adversarial network (GAN) model is constructed, consisting of a generator and a discriminator. The initial GAN model is trained using the multidimensional vector and the predefined language database to generate initial imitation voice. The initial imitation voice is compared with the user's real voice to calculate an actual similarity score. The actual similarity score is then compared with a preset standard similarity score. Based on the score differences, the multidimensional vector is adjusted to obtain a first adjustment result. User feedback on the initial imitation voice is collected through the user interface, including satisfaction scores, specific improvement suggestions, etc. Based on the user feedback results, the multidimensional vector is further adjusted to obtain a second adjustment result. The first and second adjustment results are combined to determine the final target multidimensional vector. The initial GAN model is retrained using the target multidimensional vector to generate a more accurate target imitation voice. New voice data and historical feedback are continuously collected to continuously optimize the multidimensional vector and GAN model. As needed, multiple rounds of speech synthesis and feedback processes are performed until a satisfactory imitation effect is achieved.

[0074] Specifically, the voice acquisition module combines the user's real-time and historical voice data to generate highly personalized imitation voices, meeting the user's personalized needs for voice imitation. The voice processing module converts complex voice information into multidimensional vectors to facilitate model processing and analysis. An initial GAN model is constructed using multidimensional vectors and a predefined language database, providing an efficient starting point for speech synthesis. The speech synthesis module achieves precise imitation of the user's voice through multidimensional vector combination and generative adversarial network model training. The actual similarity score approaches the preset standard, improving the accuracy of speech imitation. The voice feedback module incorporates a user feedback mechanism and iteratively optimizes the initial imitation voice based on user feedback on the initial imitation voice, ensuring that the imitation voice continuously meets user expectations and achieves continuous improvement. This system can be applied to a variety of scenarios, such as voice assistants, virtual characters, and voice conversion, improving the user experience and interactivity of these applications.

[0075] Specifically, if Figure 2 As shown, the voice processing module 20 includes:

[0076] a preprocessing unit 21 configured to perform noise elimination and echo suppression on the voice information, the audio signal, the historical voice information, and the historical audio signal to obtain a first processing result, and to segment the audio signal and the historical audio signal into a plurality of small segments to obtain a second processing result;

[0077] a feature extraction unit 22 connected to the preprocessing unit 21, configured to capture time sequence information of the first processing result to extract features of intonation and rhythm to obtain a first feature result, and further configured to extract features of timbre and pronunciation from the second processing result to obtain a second feature result;

[0078] a parameter generating unit 23 connected to the feature extracting unit 22, configured to combine the first feature result and the second feature result to obtain a multi-dimensional vector;

[0079] The model construction unit 24 is connected to the parameter generation unit 23 and is used to construct the initial generative adversarial network model using the multidimensional vector and a predefined language database as input parameters.

[0080] Specifically, multiple signals are received: current speech information, audio signals, and historical speech information and audio signals. These signals may come from different sources, such as microphones and databases. Noise suppression algorithms, such as spectral subtraction, Wiener filtering, or deep learning-based noise cancellation techniques, are applied. Background noise, such as wind and traffic noise, is reduced to improve speech signal quality. Echo suppression techniques, such as adaptive filters, are used to eliminate echo components in the signal. Echoes may be caused by reflections within the room; eliminating echoes ensures speech clarity. The signals processed with noise cancellation and echo suppression are output as input for subsequent feature extraction. The audio signal and historical audio signals are segmented into several small segments. Segmentation can be based on silence detection (detecting silence intervals between speech sounds) or fixed duration. This facilitates independent feature extraction and analysis for each segment. The segmented signal segments are output for further processing by the feature extraction unit. Time series analysis is performed on the results of the first processing to capture the temporal information of the speech. Features such as intonation (pitch changes) and rhythm (speed and pauses of speech) are extracted. The extracted time-series features are output and used to describe the dynamic changes in speech. The second processing result is then used to extract timbre and pronunciation features. This extraction method can use feature extraction techniques such as Mel-Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP). Features such as fundamental frequency (F0) can also be extracted to describe the timbre and pronunciation characteristics of speech. The extracted timbre and pronunciation features are output and used to describe the static properties of speech. The parameter generation unit combines the first feature results (time-series features) with the second feature results (timbre and pronunciation features). Vector concatenation or feature fusion techniques, such as principal component analysis (PCA), can be used. A multidimensional vector is generated, containing multidimensional features such as time-series, timbre, and pronunciation. This multidimensional vector serves as an input parameter for the model construction unit. The generated multidimensional vector is output to the model construction unit for use in constructing and training a generative adversarial network model. The model construction unit receives the multidimensional vector from the parameter generation unit and a predefined language database as input parameters. The language database may contain information such as vocabulary, grammatical rules, and phonetic annotations. Build an initial Generative Adversarial Network (GAN) model based on the input parameters. GAN structure: includes a generator and a discriminator.

[0081] Generator: Responsible for generating imitation speech, the input may be multi-dimensional vectors and other control parameters.

[0082] Discriminator: Responsible for distinguishing between real speech and generated speech. The input may be features of real speech and generated speech. The GAN model is trained using training data.

[0083] During the training process, the generator attempts to generate imitation speech that is closer and closer to real speech, while the discriminator attempts to distinguish between real speech and generated speech more and more accurately. The goal of training is to make the speech generated by the generator as close to real speech as possible in terms of auditory perception.

[0084] Specifically, the preprocessing unit improves the quality of the speech signal through noise cancellation and echo suppression, reducing noise and echo interference. The feature extraction unit extracts timbre and pronunciation features, providing a basis for generating imitation speech with the target user's timbre and pronunciation characteristics. This helps achieve highly personalized speech imitation and enhances the user experience. The parameter generation unit combines intonation, rhythm, timbre, and pronunciation features into a multidimensional vector to form a complete speech feature representation. This provides rich input information for building a generative adversarial network model, facilitating the generation of high-quality imitation speech. The model construction unit utilizes the multidimensional vector and a predefined language database to construct an initial GAN model, providing a basic framework for speech generation.

[0085] Specifically, the parameter generation unit includes:

[0086] a feature normalization subunit, configured to perform normalization processing on the first feature result and the second feature result to obtain a third processing result and a fourth processing result respectively;

[0087] A feature combination subunit is connected to the feature normalization subunit and is used to fuse the third processing result and the fourth processing result to obtain the multi-dimensional vector.

[0088] Specifically, the first feature result (intonation, rhythm features) and the second feature result (timbre, pronunciation features) from the feature extraction unit are received. The first feature result is normalized: the value range of each feature is determined (for example, by counting the maximum and minimum values). The normalization formula is applied to scale the feature values to a uniform range (for example, [0, 1] or [-1, 1]).

[0089] Normalization formula example:

[0090]

[0091] A similar normalization process is performed on the second feature result. The feature combination subunit receives the third processing result and the fourth processing result from the feature normalization subunit. The third processing result and the fourth processing result are fused to form a multidimensional vector. The fusion method can be a simple feature splicing or a more complex feature fusion technique, such as weighted fusion, principal component analysis (PCA), etc. If the third processing result is an N-dimensional vector and the fourth processing result is an M-dimensional vector, then a (N+M)-dimensional multidimensional vector can be obtained by splicing.

[0092] Feature fusion example (weighted fusion):

[0093] Assign different weights to different features and then perform weighted summation.

[0094] Weighted fusion formula example:

[0095] fusion_vector=ω1×third_result+ω2×fourth_result

[0096] Wherein, and are pre-set weights. Those skilled in the art can preset the weights according to actual conditions.

[0097] Specifically, the feature normalization unit ensures the consistency of feature vectors through normalization, facilitating stable model training. The feature combination subunit fuses multiple normalized feature results into a multidimensional vector, preserving important information from the original speech data, such as intonation, rhythm, timbre, and pronunciation, providing a comprehensive information foundation for generating high-quality imitation speech.

[0098] Specifically, the model building unit includes:

[0099] A database interface subunit, used to connect to a predefined voice database to obtain voice data;

[0100] A model construction subunit is connected to the database interface subunit and is used to perform pre-training based on the voice data and the multidimensional vector to construct the initial generative adversarial network model.

[0101] Specifically, the database interface subunit sets the database connection parameters, including the database address, port, user name, password, etc. Establish a connection with a predefined voice database. Write query statements according to requirements to extract the required voice data from the database. Voice data may include diverse voice samples in multiple languages, accents, emotions, etc. Preprocess the extracted voice data, such as format conversion and sampling rate unification, to ensure data consistency. Transfer the preprocessed voice data to the model building subunit for model training. The model building subunit receives the voice data from the database interface subunit and the multidimensional vector from the parameter generation unit. Divide the voice data into a training set, a validation set, and a test set. Select the generative adversarial network (GAN) as the basic model architecture. Set the network structure of the generator and discriminator, including the number of layers, the number of neurons in each layer, etc.

[0102] Specifically, the database interface subunit allows the model to easily access a predefined voice database, eliminating the need for manual data import and improving data usability. The model building subunit, through multi-dimensional vector input, enables the model to learn personalized voice features and generate imitation voices more suitable for the target user.

[0103] Specifically, if Figure 3 As shown, the speech synthesis module 30 includes:

[0104] A model training unit 31 is used to train the initial generative adversarial model to obtain the initial imitation speech;

[0105] a similarity calculation unit 32 connected to the model training unit 31 and configured to compare the initial imitation voice with the voice information to calculate the actual similarity score;

[0106] a score comparison unit 33 connected to the similarity comparison unit 32, for comparing the actual similarity score with a preset standard similarity score to obtain a comparison result;

[0107] The vector adjustment unit 34 is connected to the score comparison unit 33 and is used to adjust the multi-dimensional vector according to the comparison result to obtain the first adjustment result.

[0108] Specifically, the model training unit sets training parameters such as the learning rate, batch size, and number of iterations. The initial generative adversarial network model is trained using speech data and multidimensional vectors. During training, the generator attempts to generate imitation speech, while the discriminator attempts to distinguish between real speech and generated speech. The trained generative adversarial network model is output, capable of generating the initial imitation speech. The similarity calculation unit extracts features from the initial imitation speech and the source user's speech information, which may include features such as Mel-Frequency Cepstral Coefficients (MFCCs), pitch, and duration. It applies similarity calculation algorithms, such as dynamic time warping (DTW) and cosine similarity, to calculate the actual similarity score between the initial imitation speech and the source user's speech information. This actual similarity score is output for subsequent score comparison. The score comparison unit compares the actual similarity score with the standard similarity score to determine the difference between the two. The comparison result is output, indicating whether the actual similarity score meets the preset standard. The vector adjustment unit formulates an adjustment strategy for the multidimensional vector based on the comparison results. If the actual similarity score is lower than the standard similarity score, the multidimensional vector needs to be adjusted to improve the imitation effect. Fine-tuning multidimensional vectors may include adjusting feature weights, increasing or decreasing feature dimensions, etc.

[0109] Specifically, the model training unit enhances the model's generative capabilities, enabling it to better mimic the voice characteristics of a specific user. The similarity calculation unit provides a quantitative similarity score that intuitively reflects the degree of similarity between the imitated voice and the source user's voice.

[0110] Specifically, the similarity calculation unit includes:

[0111] a time alignment subunit, configured to align the initial simulated speech and the speech information in time according to dynamic time warping to obtain an alignment result;

[0112] a feature matching subunit, connected to the time alignment subunit, for performing feature matching on the alignment result to obtain a matching result;

[0113] a distance calculation subunit, connected to the matching result subunit, for calculating the Euclidean distance between the matching results to obtain a calculation result;

[0114] a similarity accumulation subunit, connected to the distance calculation subunit, for accumulating the calculation results to obtain an accumulation result;

[0115] The score normalization subunit is connected to the similarity accumulation subunit and is used to normalize the accumulation result to obtain the actual similarity score.

[0116] Specifically, the system receives the initial simulated speech and the source user's speech information. It then uses the Dynamic Time Warping (DTW) algorithm to time-align the two speech signals. The DTW algorithm overcomes the effects of different speaking speeds by finding the optimal matching path between the time series. It then outputs the aligned speech signals, ensuring temporal comparability between the two. The feature matching subunit receives the alignment results from the time alignment subunit. It extracts features, such as Mel-Frequency Cepstral Coefficients (MFCCs), from the aligned speech signals. It compares the feature vectors in corresponding time frames and calculates the degree of match between the features. It then outputs the feature matching results for subsequent distance calculation. The distance calculation subunit receives the output of the feature matching subunit. It calculates the Euclidean distance between matching feature vector pairs to quantify the difference between the features. It then outputs the Euclidean distance calculation result for each pair of feature vectors. The similarity accumulation subunit receives the output of the distance calculation subunit. It accumulates the Euclidean distance results of all feature vectors to obtain a total difference metric. This accumulated total difference metric is then output as the basis for similarity. The score normalization subunit receives the output of the similarity accumulation subunit. The accumulated results are normalized and mapped to a specific score range, such as 0 to 1 or 0 to 100. The normalized actual similarity score is output to evaluate the similarity between the imitated voice and the source user's voice.

[0117] In this embodiment, it is assumed that the initial simulated speech is 5 seconds long, has a sampling rate of 16kHz, and has a total of 80,000 sample points. The source user's speech information is 4.8 seconds long, also has a sampling rate of 16kHz, and has a total of 76,800 sample points. The speech signal is divided into frames, each of which is 25ms long and has a frame shift of 10ms. The initial simulated speech is divided into 2000 frames, and the source user's speech is divided into 1920 frames. The DTW algorithm is applied to find the optimal alignment path. An example of an alignment path: the first frame of the initial simulated speech is aligned to the first frame of the source user's speech, the second frame is aligned to the second frame, ..., the 2000th frame is aligned to the 1920th frame. After alignment, both are adjusted to 1920 frames. The alignment result is output as the aligned speech signal, with each frame still 25ms long, for a total of 1920 frames. The feature matching subunit inputs the alignment result as the aligned speech signal, with each frame of 25ms long. 13-dimensional MFCC features are extracted for each frame. A 13-dimensional MFCC feature vector is generated for each frame of the initial simulated speech and the source user's speech. The 13-dimensional MFCC feature vectors for each frame are matched. The cosine similarity between the two feature vectors is calculated. The cosine similarity ranges from -1 to 1, with 1 indicating a perfect match. For example, a cosine similarity of 0.95 for a frame indicates a high degree of feature match. The matching result matrix is 1920x1, with each element representing the cosine similarity of the corresponding frame. The distance calculation subunit inputs the matching results into the matching result matrix, which contains 1920 cosine similarity values. Euclidean distance calculation: Cosine similarity is converted to Euclidean distance. The conversion formula is: Euclidean distance = sqrt(2*(1-cosine similarity)). For example, a frame with a cosine similarity of 0.95 has a Euclidean distance of 0.22. The Euclidean distance results for each frame total 1920 distance values. The similarity accumulation subunit inputs the distance results into the 1920 Euclidean distance values. Sum up all the Euclidean distances. Example: The total difference after summing is 450.2 (hypothetical value, for example only). Total difference 450.2. Normalize the total difference to a score between 0 and 1.

[0118] Normalization formula: Actual similarity score = 1-(total difference / maximum possible difference). Calculation of maximum possible difference: Assuming the maximum difference of each frame is 1, the maximum possible difference is 1920.

[0119] Actual similarity score = 1-(450.2 / 1920)≈0.76.

[0120] The actual similarity score is 0.76, which means that the imitation voice is 76% similar to the source user's voice.

[0121] Specifically, the time alignment subunit uses the dynamic time warping (DTW) algorithm to effectively align speech signals of varying lengths, ensuring the accuracy of subsequent feature matching. The feature matching subunit distinguishes the voice characteristics of different users, enabling more personalized voice imitation. This effective feature matching algorithm reduces unnecessary computation and improves the overall system processing efficiency. The distance calculation subunit converts feature matching results into Euclidean distance, providing an intuitive and quantitative representation of similarity. The similarity accumulation subunit accumulates the matching differences across frames to comprehensively assess the similarity of the entire speech segment, preventing poor local matching from affecting the overall performance. The score normalization subunit normalizes the accumulated results to a score range of 0 to 1, ensuring a uniform and intuitively interpretable similarity score. This allows users to more easily understand the similarity score and better evaluate the effectiveness of the voice imitation.

[0122] Specifically, the scoring comparison unit includes:

[0123] a score alignment subunit, configured to determine the alignment of the actual similarity score and the standard similarity score in terms of dimension and format to obtain an alignment result;

[0124] The score calculation subunit is connected to the score alignment subunit and is used to calculate the difference between the actual similarity score and the standard similarity score according to the alignment result to obtain the comparison result.

[0125] Specifically, the standard similarity score can serve as a benchmark for evaluating the actual similarity score, helping users understand the gap between the actual similarity score and the ideal state. By comparing it with the standard similarity score, it can be determined whether the actual similarity score has achieved the expected goal or whether there is any deviation.

[0126] Specifically, the dimensions of the actual similarity score and the standard similarity score are determined, such as the number of scoring items or the type of scoring indicators. If the dimensions are inconsistent, the dimensions need to be expanded or reduced, for example, by interpolation or principal component analysis, so that the dimensions of the two scores are consistent. The formats of the actual similarity score and the standard similarity score are determined, such as the type of score value (numeric, text, etc.) and range (0-1, 0-100, etc.). If the formats are inconsistent, format conversion is required, such as converting text-based scores to numeric scores, or normalizing score values in different ranges. The score calculation subunit selects an appropriate difference calculation method based on the characteristics of the actual similarity score and the standard similarity score, for example: calculating the Euclidean distance between two score vectors, which is applicable to numeric scores. Calculating the cosine similarity between two score vectors, which is applicable to text-based or numeric scores. Calculating the Hamming distance between two score vectors, which is applicable to binary scores. Calculating the edit distance between two score sequences, which is applicable to sequence scores. The difference calculation results are weighted based on the importance of the scoring items or scoring indicators, for example, assigning greater weight to important scoring items or scoring indicators. Weights are determined using methods such as the Analytic Hierarchy Process (AHP). The calculated difference result is used as the output of the scoring comparison unit. This can be a numerical value or a vector representing the degree of difference between the actual similarity score and the standard similarity score.

[0127] In this embodiment, it is assumed that the actual similarity scores are [0.8, 0.6, 0.9] and the standard similarity scores are [0.7, 0.5, 0.8].

[0128] Dimension alignment: The dimensions of the two ratings are consistent, so there is no need to align the dimensions.

[0129] Format alignment: The two ratings have the same format and are both numerical ratings, so no format conversion is required.

[0130] Difference calculation: Calculate the distance between two rating vectors using Euclidean distance:

[0131] Difference = sqrt((0.8-0.7)^2+(0.6-0.5)^2+(0.9-0.8)^2) = 0.316

[0132] Weighted processing: Assuming that the three scoring items are of equal importance, no weighting is required.

[0133] Comparison result output: The comparison result is 0.316, indicating that the difference between the actual similarity score and the standard similarity score is 0.316.

[0134] Specifically, the score alignment subunit aligns the dimensions and format of the actual similarity score and the standard similarity score, ensuring comparability between the two scores and enabling effective comparative analysis. The score calculation subunit calculates the difference between the actual similarity score and the standard similarity score, quantifying the score comparison results for easier comparison and evaluation.

[0135] Specifically, if Figure 4 As shown, the voice feedback module 40 includes:

[0136] A feedback analysis unit 41 is used to perform text or voice analysis on the feedback result to obtain an analysis result;

[0137] a vector adjustment unit 42 connected to the feedback analysis unit 41, configured to map the analysis result to a multi-dimensional vector space to determine an adjusted vector dimension to obtain a second adjustment result;

[0138] a result integration unit 43 connected to the vector adjustment unit 42, configured to combine the first adjustment result and the second adjustment result to form a target multi-dimensional vector;

[0139] The model retraining unit 44 is connected to the result integration unit 43 and is used to prepare training data according to the target multi-dimensional vector and retrain the initial generative adversarial network model according to the prepared data to obtain the target imitation speech.

[0140] Specifically, the feedback analysis unit uses natural language processing (NLP) technology to analyze text feedback and extract information such as keywords, sentiment tendencies, and semantics. For example, sentiment analysis tools can be used to determine whether user feedback is positive, negative, or neutral. Use speech recognition technology to convert speech feedback into text, and then use NLP technology for analysis. In addition, speech feature extraction technology can be used to extract information such as the pitch, speaking speed, and volume of the speech for more in-depth analysis. The vector adjustment unit maps the analysis results to a multidimensional vector space, such as using word embedding technology to map text keywords to word vector space, and using feature extraction technology to map speech features to feature vector space. Based on the analysis results and the objective function, determine the vector dimensions that need to be adjusted. For example, if the goal is to improve the naturalness of the speech, the vector dimensions corresponding to the pitch, speaking speed, volume, and other features of the speech can be adjusted. The result integration unit merges the first adjustment result and the second adjustment result: the text analysis result and the speech analysis result are integrated to form a target multidimensional vector. The integration method can be to splice the two vectors, or to fuse them using methods such as weighted average. Model retraining unit

[0141] Prepare training data based on the target multidimensional vector and the original speech data. This training data can include speech features, text features, and the target multidimensional vector. Use the prepared training data to retrain the initial generative adversarial network model. By adjusting the model parameters, the model can generate speech that better matches the target multidimensional vector.

[0142] In this embodiment, it is assumed that the user feedback is "the voice is too mechanical" and the naturalness of the voice needs to be improved. Use sentiment analysis tools to determine that the user feedback is negative, and use speech feature extraction technology to extract information such as the pitch, speaking speed, and volume of the voice. Map "naturalness" to the word vector space, and map the speech features to the feature vector space, and determine that the vector dimensions that need to be adjusted are pitch, speaking speed, and volume. Integrate the text analysis results and the speech analysis results to form a target multidimensional vector, for example [0.8, 0.9, 0.7], which represent the adjustment targets of pitch, speaking speed, and volume, respectively. Use the target multidimensional vector and the original speech data to prepare training data, and retrain the initial generative adversarial network model so that the model can generate speech that is more consistent with the target multidimensional vector, that is, more natural speech.

[0143] Specifically, the feedback analysis unit extracts key information from user feedback and identifies user intent through text or speech analysis. The vector adjustment unit analyzes the vector dimensions corresponding to user feedback to determine the direction of model adjustment. The model retraining unit continuously collects user feedback and retrains the model, enabling continuous improvement of the model and improving speech generation quality.

[0144] Specifically, the vector adjustment unit includes:

[0145] A dimension mapping subunit, used to adjust the vector dimension according to the feedback result to obtain a mapping result;

[0146] an adjustment strategy subunit, connected to the dimension mapping subunit, for adjusting the weight of the corresponding vector dimension according to the mapping result to obtain a third adjustment result;

[0147] The vector calculation subunit is connected to the adjustment strategy subunit and is used to calculate the third adjustment result to obtain the second adjustment result.

[0148] Specifically, the dimension mapping subunit receives feedback from the feedback analysis unit. This feedback may include user evaluations of speech generation quality, specific problem descriptions, or improvement suggestions. First, the feedback is parsed to identify information related to the vector dimensions. For example, if the user feedback is "speaking too fast," this information is identified as related to the "speaking speed" dimension of the vector. The dimension relevant to the feedback is located in the multidimensional vector. For example, the vector dimension that controls speaking speed is found. Based on the feedback, adjustments to this dimension are made, such as reducing the weight of the speaking speed dimension. A mapping result is output, including the located vector dimension and its adjustment suggestion. The adjustment strategy subunit receives the mapping result from the dimension mapping subunit. Based on the mapping result, an appropriate adjustment strategy is selected. This may include weight adjustment, dimension addition, or dimension deletion. For dimensions requiring weight adjustment, specific weight adjustment values are determined. This may be based on preset rules, machine learning models, or expert experience. The selected adjustment strategy is applied to the corresponding vector dimensions. A third adjustment result, namely the multidimensional vector after weight adjustment, is output. The vector calculation subunit receives the third adjustment result from the adjustment strategy subunit. Based on the adjusted weights, the components of the multidimensional vector are recalculated. The reconstructed vector is normalized to ensure that it meets a specific format or standard. Verification is performed to ensure that the adjusted vector meets the expected quality requirements or user needs. The result of the second adjustment, the adjusted multidimensional vector, is output to guide improvements to the speech generation model.

[0149] Specifically, the Dimension Mapping subunit accurately identifies and locates the vector dimensions that require adjustment based on feedback, ensuring targeted adjustments. The Adjustment Strategy subunit effectively optimizes speech generation quality by adjusting the weights of vector dimensions, ensuring that the generated speech better meets user needs. The Vector Calculation subunit normalizes the results of the third processing step to ensure that the adjusted vectors maintain consistent format with other vectors, facilitating subsequent processing.

[0150] Specifically, the model retraining unit includes:

[0151] A model configuration subunit, configured to adjust the parameters of the initial generative adversarial network model according to new training data and objectives to obtain a configuration result;

[0152] The training execution subunit is connected to the model configuration subunit and is used for performing training according to the configuration result to obtain the target imitation speech.

[0153] Specifically, the model configuration subunit receives the target multidimensional vector from the result integration unit, which contains feature information adjusted based on user feedback. It also receives new training data, which may be supplemented original speech data, adjusted speech data, or simulated speech data. Parameter adjustment: Based on the target multidimensional vector, the parameters of the initial generative adversarial network (GAN) model are adjusted. This includes, but is not limited to, the weights, learning rate, and loss function of the generator and decomposer. If necessary, the model architecture is optimized, such as adding or removing network layers or changing activation functions, to better adapt to the new training data. The adjusted model configuration is verified to ensure its correctness and effectiveness. The configuration results, including the adjusted model parameters and structure information, are output to the training execution subunit. The training execution subunit receives the configuration results from the model configuration subunit. Based on the configuration results, training data is prepared. This may include preprocessing steps such as data augmentation, normalization, and segmentation. Model training is performed using the prepared training data according to the configuration results. This includes steps such as forward propagation, loss calculation, backpropagation, and parameter update. During training, the model's performance, such as loss and accuracy, is monitored in real time to ensure training effectiveness. Based on the performance monitoring results, adjustments to the training process are made as necessary, such as adjusting the learning rate or increasing the number of training rounds. The trained target imitation speech is output, which is generated based on the new training data and the adjusted model to better meet user needs.

[0154] In this embodiment, based on the target multidimensional vector, the speech rate control parameter in the generator is reduced by 10% to adapt to the user's feedback on the speech rate. The new training data is normalized and the data set is divided into a training set and a validation set. The model is trained using the training set data, and the model performance is evaluated using the validation set data after each round of training. If it is found that the loss value does not decrease significantly after several consecutive rounds of training, the learning rate is reduced by half and training continues. Through such a detailed method, the model retraining unit can effectively adjust the model according to the new training data and targets, and generate a target imitation speech that better meets the user's needs.

[0155] Specifically, the model configuration subunit can personalize the parameters of the initial generative adversarial network model based on new training data and user feedback targets, making the model more tailored to the needs of specific users. The model retraining unit significantly improves the quality of generated speech through the collaborative work of configuration and training, making it more natural and more in line with user needs.

[0156] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

[0157] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. An AI-based voice imitation system, characterized in that: include: The voice acquisition module is used to obtain the user's audio signal and voice information, and also to obtain the user's historical voice information and historical audio signal; a speech processing module, connected to the speech acquisition module, for combining the speech information, the audio signal, the historical speech information, and the historical audio signal to obtain a multidimensional vector, and constructing an initial generative adversarial network model with the multidimensional vector and a predefined language database; a speech synthesis module, connected to the speech processing module, configured to train the initial generative adversarial network model to obtain an initial simulated speech, compare the initial simulated speech with the speech information to obtain an actual similarity score, and compare the actual similarity score with a preset standard similarity score to adjust the multidimensional vector to obtain a first adjustment result; A speech feedback module is connected to the speech synthesis module and is used to obtain a feedback result based on the user's feedback on the initial imitation speech, adjust the multidimensional vector based on the feedback result to obtain a second adjustment result, merge the first adjustment result and the second adjustment result to obtain a target multidimensional vector, and retrain the initial generative adversarial model based on the target multidimensional vector to obtain a target imitation speech.

2. The AI-based voice imitation system according to claim 1, characterized in that: The speech processing module includes: a preprocessing unit, configured to perform noise elimination and echo suppression on the voice information, the audio signal, the historical voice information, and the historical audio signal to obtain a first processing result, and to segment the audio signal and the historical audio signal into a plurality of small segments to obtain a second processing result; a feature extraction unit connected to the preprocessing unit, configured to capture time sequence information of the first processing result to extract features of intonation and rhythm to obtain a first feature result, and further configured to extract features of timbre and pronunciation from the second processing result to obtain a second feature result; a parameter generating unit, connected to the feature extracting unit, and configured to combine the first feature result and the second feature result to obtain the multidimensional vector; A model construction unit is connected to the parameter generation unit and is used to construct the initial generative adversarial network model using the multidimensional vector and a predefined language database as input parameters.

3. The AI-based voice imitation system according to claim 2, characterized in that: The parameter generation unit includes: a feature normalization subunit, configured to perform normalization processing on the first feature result and the second feature result to obtain a third processing result and a fourth processing result respectively; A feature combination subunit is connected to the feature normalization subunit and is used to fuse the third processing result and the fourth processing result to obtain the multi-dimensional vector.

4. The AI-based voice imitation system according to claim 3, characterized in that: The model building unit includes: A database interface subunit, used to connect to a predefined voice database to obtain voice data; A model construction subunit is connected to the database interface subunit and is used to perform pre-training based on the voice data and the multidimensional vector to construct the initial generative adversarial network model.

5. The AI-based voice imitation system according to claim 4, characterized in that: The speech synthesis module includes: A model training unit, configured to train the initial generative adversarial model to obtain the initial imitation speech; a similarity calculation unit, connected to the model training unit, for comparing the initial simulated voice with the voice information to calculate the actual similarity score; a score comparison unit, connected to the similarity comparison unit, for comparing the actual similarity score with a preset standard similarity score to obtain a comparison result; A vector adjustment unit is connected to the score comparison unit and is used to adjust the multi-dimensional vector according to the comparison result to obtain the first adjustment result.

6. The AI-based voice imitation system according to claim 5, characterized in that: The similarity calculation unit includes: a time alignment subunit, configured to align the initial simulated speech and the speech information in time according to dynamic time warping to obtain an alignment result; a feature matching subunit, connected to the time alignment subunit, for performing feature matching on the alignment result to obtain a matching result; a distance calculation subunit, connected to the matching result subunit, for calculating the Euclidean distance between the matching results to obtain a calculation result; a similarity accumulation subunit, connected to the distance calculation subunit, for accumulating the calculation results to obtain an accumulation result; The score normalization subunit is connected to the similarity accumulation subunit and is used to normalize the accumulation result to obtain the actual similarity score.

7. The AI-based voice imitation system according to claim 6, characterized in that: The scoring comparison unit includes: a score alignment subunit, configured to determine the alignment of the actual similarity score and the standard similarity score in terms of dimension and format to obtain an alignment result; The score calculation subunit is connected to the score alignment subunit and is used to calculate the difference between the actual similarity score and the standard similarity score according to the alignment result to obtain the comparison result.

8. The AI-based voice imitation system according to claim 7, characterized in that: The voice feedback module includes: A feedback analysis unit, configured to perform text or voice analysis on the feedback result to obtain an analysis result; a vector adjustment unit, connected to the feedback analysis unit, configured to map the analysis result to a multi-dimensional vector space to determine an adjusted vector dimension to obtain a second adjustment result; a result integration unit connected to the vector adjustment unit, configured to combine the first adjustment result and the second adjustment result to form a target multi-dimensional vector; A model retraining unit is connected to the result integration unit and is used to prepare training data according to the target multidimensional vector and retrain the initial generative adversarial network model according to the prepared data to obtain the target imitation speech.

9. The AI-based voice imitation system according to claim 8, characterized in that: The vector adjustment unit includes: A dimension mapping subunit, used to adjust the vector dimension according to the feedback result to obtain a mapping result; an adjustment strategy subunit, connected to the dimension mapping subunit, for adjusting the weight of the corresponding vector dimension according to the mapping result to obtain a third adjustment result; The vector calculation subunit is connected to the adjustment strategy subunit and is used to calculate the third adjustment result to obtain the second adjustment result.

10. The AI-based voice imitation system according to claim 9, characterized in that: The model retraining unit includes: A model configuration subunit, configured to adjust the parameters of the initial generative adversarial network model according to new training data and objectives to obtain a configuration result; The training execution subunit is connected to the model configuration subunit and is used for performing training according to the configuration result to obtain the target imitation speech.

Images