A method and system for optimizing speaker separation based on unsupervised

Abstract

The present invention discloses an unsupervised speaker separation optimization method and system, which relates to the technical field of speaker separation, including collecting voice data and preprocessing the voice data; extracting and classifying features through unsupervised data annotation; constructing an unsupervised reinforcement learning strategy optimization algorithm, optimizing the model, and using unprocessed voice data to test and verify the model. The method of the present invention extracts and classifies features through unsupervised data annotation, realizes automatic recognition and classification of voice features without manual annotation, reduces annotation costs, and the model can adapt to unknown speakers, thereby enhancing the generalization ability and robustness of the model; by constructing an unsupervised reinforcement learning strategy optimization algorithm, the model can better handle noise and changes in practical applications, improves the accuracy and adaptability of the model, and uses unprocessed voice data to test and verify the model, thereby verifying the generalization ability and practicality of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of speaker separation, and in particular to an unsupervised speaker separation optimization method and system. Background Art

[0002] Speaker separation technology has broad application prospects in voice communication, speech recognition, audio processing and other fields. With the development of artificial intelligence technology, the application effect of speaker separation technology has been significantly improved. However, the existing speaker separation technology still faces some challenges and limitations. At present, speaker separation technology mainly relies on supervised learning methods. Supervised learning methods usually require a large amount of labeled data for model training. The cost of obtaining labeled data is high and the process is cumbersome. Not only does it require a lot of manpower and material resources for manual labeling, but also due to the complexity of the labeling work itself, the labeling quality is difficult to guarantee. The high cost and low efficiency of the labeling process have become an important factor restricting the development of speaker separation technology. In addition, due to the limited diversity of training data, the generalization ability of existing speaker separation models is limited, and it is difficult to deal with the speech signals of unknown speakers in specific scenarios, resulting in unsatisfactory performance in practical applications. Especially when facing complex acoustic environments and changeable speakers, the performance of the model will be significantly reduced.

[0003] Based on the above technical problems, the unsupervised speaker separation optimization method proposed in the present invention, firstly, through the unsupervised learning strategy, can automatically label the data without manual intervention, which can reduce the cost of data preparation and make the speaker separation technology more efficient and practical. Secondly, in terms of enhancing the generalization ability, the present invention optimizes and adjusts the speaker separation model through the unsupervised extraction of feature vectors, so that the model can better adapt to the speech signals of unknown speakers in specific scenarios and improve the generalization ability. Compared with the prior art, the advantages of the present invention are: on the one hand, it reduces the need for manual labeling of data and reduces the cost; on the other hand, it improves the accuracy and robustness of the speaker separation model, so that it has better performance in practical applications. In short, the present invention brings an innovative solution to the field of speaker separation technology and is expected to promote the development of this field. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: the existing speaker separation technical methods have high data labeling costs, low labeling efficiency, weak model generalization ability, and how to automatically label data without manual intervention and improve the model generalization ability.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a speaker separation optimization method based on unsupervised speech, comprising collecting speech data and preprocessing the speech data; extracting and classifying features through unsupervised data annotation; constructing an unsupervised reinforcement learning strategy optimization algorithm, optimizing the model, and using unprocessed speech data to test and verify the model.

[0007] As a preferred solution of the unsupervised speaker separation optimization method described in the present invention, the collecting of speech data includes collecting speech data of application scenarios based on the speaker separation model application scenarios.

[0008] As a preferred solution of the unsupervised speaker separation optimization method described in the present invention, the preprocessing includes collecting speech data, removing invalid data, unifying the number of audio file channels and adding environmental noise.

[0009] As a preferred solution of the unsupervised speaker separation optimization method described in the present invention, the unsupervised data annotation includes automatically identifying vocal segments through voice activity detection technology enhanced by pitch sensitivity, completing speaker change and confirmation using voice individual feature extraction technology enhanced by time series analysis, and completing automatic classification of voice individual features and role classification through voice individual classification technology driven by Euclidean distance.

[0010] As a preferred solution of the unsupervised speaker separation optimization method described in the present invention, the feature extraction and classification includes generating a partial sequence of speech activity based on a speech activity detection technology enhanced by pitch sensitivity; extracting a speech individual feature sequence containing time series information through a speech individual feature extraction technology enhanced by time series analysis based on the partial sequence of speech activity; and generating a matrix containing speech individual feature space information and a speech individual classification sequence through a speech individual classification technology driven by Euclidean distance based on the speech individual feature sequence containing time series information.

[0011] As a preferred solution of the unsupervised speaker separation optimization method described in the present invention, the optimization model includes constructing a speaker separation pre-training model for the application scenario of the speaker separation model of an unsupervised reinforcement learning strategy optimization algorithm based on unsupervised data labeling for secondary training.

[0012] As a preferred solution of the unsupervised speaker separation optimization method described in the present invention, the model testing and verification includes using unprocessed speech data, inputting the optimized speaker separation model, performing speaker separation, and using the detection error rate to evaluate the model.

[0013] Another object of the present invention is to provide an unsupervised speaker separation optimization system that can perform feature extraction and classification through unsupervised data annotation, thereby solving the problem of high data annotation cost and low annotation efficiency in current speaker separation technology.

[0014] As a preferred solution of the unsupervised speaker separation optimization system described in the present invention, it includes: a speech data preprocessing module, an unsupervised data labeling module, and a model optimization module; the speech data preprocessing module is used to collect speech data of speaker separation model application scenarios and preprocess the speech data; the unsupervised data labeling module is used to perform feature extraction and classification based on the preprocessed speech data through unsupervised data labeling; the model optimization module is used to construct an unsupervised reinforcement learning strategy optimization algorithm, optimize the speaker separation model, and perform model testing and verification.

[0015] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of an unsupervised speaker separation optimization method.

[0016] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of an unsupervised speaker separation optimization method.

[0017] Beneficial effects of the present invention: the unsupervised speaker separation optimization method provided by the present invention collects voice data and pre-processes the voice data, thereby realizing the cleaning and format unification of the original data, ensuring the data quality of subsequent processing, providing reliable basic data for feature extraction and model training, and achieving the goal of reducing noise interference and improving data consistency; feature extraction and classification are performed through unsupervised data annotation, thereby realizing automatic recognition and classification of voice features without manual annotation, reducing the annotation cost, and improving the automation level of the system, so that the model can adapt to unknown speakers, thereby achieving enhanced model generalization ability and robustness; by constructing an unsupervised reinforcement learning strategy optimization algorithm, the refined adjustment of the interlocutor separation model is realized, the model can better handle noise and changes in practical applications, and the accuracy and adaptability of the model are improved; the unprocessed voice data is used for model testing and verification, thereby ensuring that the optimized model can maintain high performance in practical applications, and achieving the goal of verifying the generalization ability and practicality of the model. The present invention achieves better results in terms of accuracy, robustness and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0019] Figure 1 An overall flow chart of an unsupervised speaker separation optimization method provided for the first embodiment of the present invention;

[0020] Figure 2 A flow chart of an unsupervised data labeling and unsupervised reinforcement learning strategy optimization algorithm based on an unsupervised speaker separation optimization method provided in the first embodiment of the present invention;

[0021] Figure 3 A schematic diagram of modules of an unsupervised speaker separation optimization system provided in the third embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned purposes, 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 drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0023] Example 1, reference Figure 1-Figure 2 , is an embodiment of the present invention, and provides an unsupervised speaker separation optimization method, comprising:

[0024] S1: Collect voice data and pre-process the voice data.

[0025] Furthermore, collecting voice data includes collecting application scenario voice data based on the speaker separation model application scenario.

[0026] It should be noted that the preprocessing includes collecting speech data, removing invalid data, unifying the number of channels of audio files and adding environmental noise.

[0027] It should also be noted that based on the application scenarios of the speaker separation model, voice data of the application scenarios are collected. Taking the PyanNet model as an example, this model is trained based on conversational voice data in various English scenarios. When used in Chinese scenarios for speaker separation tasks, a large amount of outbound call recording data must be collected first. The data should cover different speakers, background noise, and call environments. At the same time, the data set contains diverse voice samples, such as speakers of different genders, ages, accents, and languages, to improve the generalization and robustness of the model.

[0028] It should also be noted that data preprocessing removes invalid data, such as deleting audio clips that are too short. Audio clips that are too short cannot provide enough feature information, which may cause deviations in the model's feature learning and lead to a decline in model performance. At the same time, the number of audio file channels is unified. The task scenario of speaker separation usually requires single-channel recording, so the audio file needs to be converted into single-channel audio. At the same time, simulated environmental noise is added to enhance the data set, so that the model can better adapt to real application scenarios and improve the robustness of the model.

[0029] It should also be noted that the collection of speech data for application scenarios ensures the pertinence and practicality of the collected data, so that the speaker separation model can be optimized for specific application scenarios, and real and diverse training samples can be provided for the model, thereby improving the performance and adaptability of the model in practical applications. By removing invalid or low-quality speech fragments, the overall quality of the data set is improved, ensuring the efficiency and accuracy of model training, unifying the number of audio file channels, and standardizing the data format to facilitate unified processing and analysis of the model, eliminating the processing complexity caused by channel differences, and improving the stability and repeatability of model training. By adding simulated environmental noise to the training data, the robustness of the model is enhanced, enabling the model to better adapt to various noise environments in practical applications, improving the performance of the model in complex acoustic environments, and ensuring its reliability in practical applications.

[0030] S2: Feature extraction and classification through unsupervised data annotation.

[0031] Furthermore, unsupervised data annotation includes automatic recognition of vocal segments through voice activity detection technology enhanced by pitch sensitivity, speaker change and confirmation through voice individual feature extraction technology enhanced by time series analysis, and automatic classification and role classification of voice individual features through voice individual classification technology driven by Euclidean distance.

[0032] It should be noted that feature extraction and classification include speech activity detection technology based on pitch sensitivity enhancement to generate a partial sequence of speech activity; based on the partial sequence of speech activity, a speech individual feature extraction technology enhanced by time series analysis is used to extract a speech individual feature sequence containing time series information; based on the speech individual feature sequence containing time series information, a speech individual classification technology driven by Euclidean distance is used to generate a matrix containing speech individual feature space information and a speech individual classification sequence.

[0033] It should also be noted that the voice activity detection technology based on pitch sensitivity enhancement generates a partial sequence of voice activities, which is expressed as:

[0034]

[0035] Among them, x t is the input speech signal, P is the data preprocessing to reduce noise and echo of the input speech signal, E is the pitch sensitivity enhancement, F is the extraction feature, y t is the comparison result between the feature value and the threshold T, Y is the output speech activity part sequence, otherwise is the non-speech activity part, and t is the time point.

[0036] Through the speech individual feature extraction technology enhanced by time series analysis, the speech individual feature sequence containing time series information is extracted, which is expressed as:

[0037] v t =F((P(w t )|f i (w))),V={v1,v2,...,v t}

[0038] Among them, v t is the individual feature of speech, w t is the speech signal divided by the dynamic time-series speech feature window, w is the entire input audio signal, and f i In order to dynamically set the window size according to the speech characteristics, V is a sequence of individual speech features containing time sequence information.

[0039] Through the speech individual classification technology driven by Euclidean distance, a matrix containing speech individual feature space information and a speech individual classification sequence are generated, which can be expressed as:

[0040] V={v1,v2,...,v n}

[0041]

[0042] S={s1,s2,...,s n}

[0043] C = f DensityClustering (S)

[0044] o n = argmax(C),O = {o1,o2,...,o n}

[0045] Among them, v n 、v j are the nth and jth individual speech features in the sequence, is the Euclidean distance between the nth individual voice feature and the jth individual voice feature, d euclidean is the Euclidean distance calculation, m is the dimension of the individual speech feature, k is the two-dimensional vector of the individual speech feature, s n is the Euclidean distance sequence between the nth individual voice feature and other individual voice features, S is an n×n matrix, C is the clustering result, f DensityClustering is the density clustering algorithm, o n is the individual voice feature v n role classification, O is the individual speech classification sequence.

[0046] It should also be noted that Figure 2The algorithm flow chart for unsupervised data labeling and unsupervised reinforcement learning strategy optimization is provided. The voice activity detection technology based on pitch sensitivity enhancement ensures that the data to be processed subsequently only contains meaningful human voice signals. The collected unlabeled voice data is processed to automatically detect the segments containing human voice from the original voice signal while excluding non-voice signals. The voice activity detection technology with pitch sensitivity enhancement solves the problem that traditional detection methods may fail when the signal-to-noise ratio is low, so as to provide more reliable detection results. After filtering the non-voice signals, the voice activity detection needs to change and confirm the speaker of the active voice part. In order to ensure more accurate distinction between speaker changes and confirmations, the voice individual extraction technology enhanced by time series analysis is introduced, and a dynamic time series voice feature window of size i is introduced to accurately segment the key parts of the voice activity. While capturing the instantaneous features of the voice, the dynamic changes of the voice signal on the time axis are analyzed, so as to extract richer and more distinctive time series voice individual features. When using the voice enhanced by time series analysis, the voice After the individual feature extraction technology is used, a time-series speech individual feature sequence can be obtained, which contains the speech individual feature information of each window. After obtaining the speech individual feature sequence, each speech individual feature in the sequence is classified to distinguish the speech activity parts of different speakers in the audio. The speech individual classification technology driven by Euclidean distance can automatically classify each speech individual feature to the corresponding speaker role under unlabeled conditions, and accurately calculate the similarity of each feature point in the speech individual feature sequence. The Euclidean distance is used to calculate the similarity between each speech individual feature and all other speech individual features in the sequence. Each speech individual feature is mapped into a matrix, which contains the similarity between speech individual features, relative positions in the feature space and spatial relationships. Finally, all matrices are spliced ​​as input to the density clustering algorithm to realize automatic classification of speech individual features. Finally, the speech individual classification algorithm will output a sequence, which contains the speech individual features and role classification of each audio segment.

[0047] It should also be noted that the feature extraction and classification process of unsupervised data annotation, through the voice activity detection technology enhanced by pitch sensitivity, automatically identifies human voice fragments, solves the problem that traditional detection methods may fail when the signal-to-noise ratio is low, improves the accuracy and reliability of voice activity detection, and provides a high-quality data foundation for subsequent feature extraction and classification. Through the voice individual feature extraction technology enhanced by time series analysis, richer and more distinctive time series features are extracted, which solves the limitations of traditional feature extraction methods in distinguishing different speakers, improves the distinctiveness and expressiveness of individual voice features, and provides more accurate and reliable feature information for voice individual classification. Through the voice individual classification technology driven by Euclidean distance, automatic classification and role classification of individual voice features are realized, which solves the problem of how to accurately distinguish different speakers under unlabeled conditions, improves the accuracy and automation of individual voice classification, and provides effective data support for the training and application of speaker separation models.

[0048] S3: Build an unsupervised reinforcement learning strategy optimization algorithm, optimize the model, and use unprocessed speech data for model testing and verification.

[0049] Furthermore, the optimization model includes constructing an unsupervised reinforcement learning strategy optimization algorithm for the speaker separation model application scenario based on unsupervised data annotation and performing secondary training on the speaker separation pre-training model.

[0050] It should be noted that model testing and verification includes using unprocessed speech data, inputting the optimized speaker separation model, performing speaker separation, and using the detection error rate to evaluate the model.

[0051] It should also be noted that the unsupervised reinforcement learning strategy optimization algorithm is constructed as follows:

[0052] M optimized =f UE (f FeatureFusion (f FineTune (M,O),S))

[0053] Where M is the speaker separation pre-training model, f FeatureFusion For feature extraction and learning of secondary training, f FeatureFusion is feature fusion, f UE For unsupervised reinforcement learning, M optimized This is the optimized speaker separation model.

[0054] It should also be noted that Figure 2This is a flow chart of the unsupervised data labeling and unsupervised reinforcement learning strategy optimization algorithm. After completing the individual speech classification, the speaker separation model of the application scenario is optimized, and the speaker separation pre-training model of the speech individual classification output sequence is used for secondary training. By learning the data of the actual application scenario, the speaker separation model can be made more versatile and robust. At the same time, in order to avoid overfitting in the role classification output by the Euclidean distance-driven speech individual classification technology in the unsupervised data labeling during the speaker classification model optimization process, the matrix features output by the density clustering algorithm in the unsupervised data labeling are integrated while the model learns the data features. Through the matrix features, the model can learn the spatial distribution of new data and guide the model to pay more attention to the spatial distribution in the data. After optimizing the speaker separation pre-training model using the unsupervised reinforcement learning strategy optimization algorithm, the robustness and accuracy of the speaker separation model in the actual application scenario data can be improved.

[0055] It should also be noted that the unprocessed speech data is input into the optimized speaker separation model for speaker separation, and the detection error rate is used for model evaluation, which is expressed as:

[0056]

[0057] Among them, DER is the detection error rate, FA is the model mistakenly identifying the non-speech area as the speech area of ​​a certain speaker, MD is the model failing to identify the speech area of ​​a certain speaker, SC is the model mistakenly classifying the speech of one speaker as the speech area of ​​another speaker, and TDA is the total duration of all audio.

[0058] It should also be noted that after completing the tuning of the speaker separation model, in order to ensure that the model performs as expected in actual applications and can work stably under various conditions, the model is tested and verified, and a batch of audio data that did not participate in the model training is collected from the speaker separation model application scenario. The data covers different speakers, background noise and recording environments. The test data is used to separate the speakers through the optimized speaker separation model, and the detection error rate is used to evaluate the model. The detection error rate can reflect the accuracy and reliability of the optimized speaker separation model in actual applications.

[0059] It should also be noted that the speaker separation model is optimized through unsupervised data labeling and feature extraction, combined with unsupervised reinforcement learning strategies, which improves the generalization ability and adaptability of the model, and provides a more efficient and reliable training and optimization method for the speaker separation model. Through secondary training and feature fusion, the performance and accuracy of the model are further improved, and the accuracy and reliability of the model in practical applications are improved, ensuring that the speaker separation model can adapt to different application scenarios and data sets. The optimized model is tested and verified using unprocessed speech data to evaluate the performance and generalization ability of the model, verify the accuracy and reliability of the model, and ensure that the performance of the model in practical applications meets expectations and works stably under various conditions.

[0060] Example 2 is an embodiment of the present invention, which provides an unsupervised speaker separation optimization method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0061] First, for the outbound call application scenario of a company using the speaker separation model, a speech data set totaling 200 hours was collected, including multiple speakers, different background noises, and recording environments. After data collection, the following preprocessing was performed to remove invalid speech fragments, unify the sampling rate of the audio files to 44.1kHz and the number of channels to stereo, and add environmental noise with different signal-to-noise ratios (SNR) to the audio to simulate the real application environment.

[0062] Next, unsupervised data labeling technology is used to extract and classify features of speech data. First, the voice clips are automatically identified through the pitch-sensitive enhanced speech activity detection technology, and about 15% of non-vocal noise is removed. Then, the voice individual feature extraction technology enhanced by time series analysis is used to extract features from the voice clips, and the change points of the speaker are successfully identified and marked. Finally, based on the Euclidean distance-driven speech individual classification technology, the automatic classification of speech individual features is realized, forming a highly accurate speech individual classification sequence.

[0063] During the stage of constructing the unsupervised reinforcement learning strategy optimization algorithm, the speaker separation pre-training model for the speaker separation model application scenario was trained a second time. By integrating the matrix features output by the density clustering algorithm in unsupervised data labeling, the model can learn the spatial distribution of new data while learning the data features, thereby improving the robustness and accuracy of the model in actual application scenario data.

[0064] During the model testing and verification phase, unprocessed speech data was input into the optimized speaker separation model to perform speaker separation, and the model was evaluated using the detection error rate. The test dataset contained 50 hours of speech data, which was not involved in the model training to ensure the objectivity of the evaluation results.

[0065] Refer to Table 1 to record and analyze the experimental data.

[0066] Table 1 Experimental data record table

[0067]

[0068]

[0069] By comparing the detection error rates before and after optimization, it can be clearly seen that the accuracy of the optimized speaker separation model has been improved. In all test samples, the detection error rate of the optimized model has been reduced by 0.08 on average, which shows that the model has enhanced its performance in identifying the speaker's speech area and avoiding misclassification. Especially under low signal-to-noise ratio (SNR) conditions, the performance of the optimized model has been significantly improved. For example, under the condition of SNR of -5dB, the detection error rate of the original model is 0.25, while the detection error rate of the optimized model is reduced to 0.12, which shows that the optimization strategy effectively improves the robustness of the model in noisy environments. In addition, for The optimized model also shows good performance for speech samples containing more speakers. In samples containing 4 speakers, the detection error rate of the optimized model decreased by 0.09 on average, indicating that the accuracy and reliability of the model are improved when dealing with complex multi-speaker scenarios. In summary, the application of the unsupervised reinforcement learning strategy optimization algorithm proposed in this study in the speaker separation model not only improves the accuracy of the model, but also enhances the versatility of the model for different recording environments and numbers of speakers. Compared with the prior art, the present invention shows innovation and advantages in reducing the detection error rate and improving the robustness of the model. Therefore, our invention is creative.

[0070] Example 3, reference Figure 3 , which is an embodiment of the present invention, provides an unsupervised speaker separation optimization system, including a speech data preprocessing module, an unsupervised data labeling module, and a model optimization module.

[0071] The speech data preprocessing module is used to collect speech data for speaker separation model application scenarios and to preprocess the speech data; the unsupervised data labeling module is used to perform feature extraction and classification based on the preprocessed speech data through unsupervised data labeling; the model optimization module is used to construct an unsupervised reinforcement learning strategy optimization algorithm, optimize the speaker separation model, and perform model testing and verification.

[0072] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0073] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0074] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0075] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An unsupervised speaker separation optimization method, characterized in that: include: Collect voice data and pre-process the voice data; Feature extraction and classification through unsupervised data annotation; Build an unsupervised reinforcement learning strategy optimization algorithm, optimize the model, and use unprocessed speech data to test and verify the model; The unsupervised data annotation includes automatically identifying human voice segments through voice activity detection technology enhanced by pitch sensitivity, completing speaker change and confirmation using voice individual feature extraction technology enhanced by time series analysis, and completing automatic classification of voice individual features and role classification through voice individual classification technology driven by Euclidean distance; The feature extraction and classification includes a voice activity detection technique based on pitch sensitivity enhancement to generate a partial sequence of voice activities, expressed as: Among them, x t is the input speech signal, P is the data preprocessing to reduce noise and echo of the input speech signal, E is the pitch sensitivity enhancement, F is the extraction feature, y t is the comparison result between the feature value and the threshold T, Y is the output speech activity part sequence, otherwise is the non-speech activity part, and t is the time point; Based on the partial sequence of speech activity, the speech individual feature extraction technology enhanced by time sequence analysis is used to extract the speech individual feature sequence containing time sequence information, which is expressed as: v t =F((P(w t )|f i (w))),V={v1,v2,...,v t } Among them, v t is the individual feature of speech, w t is the speech signal divided by the dynamic time-series speech feature window, w is the entire input audio signal, and f i To dynamically set the window size according to the speech characteristics, V is the individual speech feature sequence containing time sequence information; Based on the speech individual feature sequence containing time sequence information, the speech individual classification technology driven by Euclidean distance is used to generate a matrix containing speech individual feature space information and a speech individual classification sequence, which is expressed as: V={v1,v2,...,v n } S={s1,s2,...,s n } C=f DensityClustering (S) o n =argmax(C),O={o1,o2,...,o n } Among them, v n 、v j are the nth and jth individual speech features in the sequence, is the Euclidean distance between the nth individual voice feature and the jth individual voice feature, d euclidean is the Euclidean distance calculation, m is the dimension of the individual speech feature, k is the two-dimensional vector of the individual speech feature, s n is the Euclidean distance sequence between the nth individual voice feature and other individual voice features, S is an n×n matrix, C is the clustering result, f DensityClustering is the density clustering algorithm, o n is the individual voice feature v n , O is the classification sequence of individual voices; The optimization model includes constructing an unsupervised reinforcement learning strategy optimization algorithm interlocutor separation model application scenario based on unsupervised data annotation and performing secondary training on the interlocutor separation pre-training model; Construct an unsupervised reinforcement learning strategy optimization algorithm, expressed as: M optimized =f UE (f FeatureFusion (F fineTune (M,O),S)) Among them, M is the speaker separation pre-training model, F FineTune For feature extraction and learning of secondary training, f FeatureFusion is feature fusion, f UE For unsupervised reinforcement learning, M optimized This is the optimized speaker separation model.

2. The unsupervised speaker separation optimization method according to claim 1, characterized in that: The collecting of voice data includes collecting application scenario voice data based on the speaker separation model application scenario.

3. The unsupervised speaker separation optimization method according to claim 2, characterized in that: The preprocessing includes collecting voice data, removing invalid data, unifying the number of audio file channels and adding environmental noise.

4. The unsupervised speaker separation optimization method according to claim 3, characterized in that: The model testing and verification includes using unprocessed speech data, inputting into an optimized speaker separation model, performing speaker separation, and using a detection error rate to perform model evaluation.

5. A system using the unsupervised speaker separation optimization method according to any one of claims 1 to 4, characterized in that: Including speech data preprocessing module, unsupervised data labeling module, and model optimization module; The speech data preprocessing module is used to collect speech data of the speaker separation model application scenario and preprocess the speech data; The unsupervised data annotation module is used to perform feature extraction and classification based on the preprocessed speech data through unsupervised data annotation; The model optimization module is used to construct an unsupervised reinforcement learning strategy optimization algorithm, optimize the speaker separation model, and perform model testing and verification.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the unsupervised speaker separation optimization method according to any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the unsupervised speaker separation optimization method according to any one of claims 1 to 4 are implemented.

Images