A voiceprint retrieval method, device, electronic device and storage medium

Abstract

This application provides a voiceprint retrieval method, apparatus, electronic device, and storage medium. The method includes: firstly, segmenting the sound to be identified into M segments using a preset segmentation strategy, and extracting features from these segments and the original sound to be identified; secondly, constructing M+1 voiceprint models based on the extracted features; thirdly, performing independent voiceprint retrieval on each voiceprint model to generate M+1 sets of top N candidate ranking results, each set containing the N target voiceprints with the highest matching degree in the voiceprint database and their matching degree; finally, updating the top N candidate ranking of each voiceprint model by comprehensively comparing the matching degree of the same target voiceprints in each set of results, so as to output a more accurate and reliable ranking result. This application can improve the accuracy of voiceprint retrieval.

Description

Technical Field

[0001] The present application relates to the field of audio processing, and in particular to a voiceprint retrieval method, device, electronic device and storage medium. Background Art

[0002] Voiceprint recognition technology is a type of biometric technology. It is a technology that automatically identifies the speaker's identity based on the speaker information contained in the voice waveform. Voiceprint retrieval is the process of quickly finding matching or similar voiceprint records in a large amount of voiceprint data through voiceprint recognition technology and outputting the topN candidate rankings.

[0003] When voiceprint retrieval compares the voiceprint model of the speech to be recognized with the voiceprint data in the built database, the entire speech to be recognized is subjected to feature extraction and voiceprint modeling, and then the model is compared and scored with the voice model in the built database to find matching or similar voiceprint records and output the topN candidate ranking. However, when the speech to be recognized is long, modeling the entire segment will result in the loss of some speech features, affecting the accuracy of the topN candidate ranking of the retrieval result similarity. Summary of the invention

[0004] In view of this, the embodiments of the present application provide a voiceprint retrieval method, device, electronic device and storage medium, which can improve the accuracy of voiceprint retrieval.

[0005] The technical solution of the embodiment of the present application is implemented as follows:

[0006] In a first aspect, an embodiment of the present application provides a voiceprint retrieval method, the method comprising:

[0007] Segment the sound to be recognized by using a preset segmentation strategy to obtain M sound segments to be recognized, where M is a positive integer greater than or equal to 2;

[0008] Performing feature extraction on the sound to be identified and the M sound segments to be identified respectively, and constructing M+1 voiceprint models based on the extraction results of the feature extraction;

[0009] Perform voiceprint retrieval on each of the M+1 voiceprint models respectively to obtain M+1 groups of topN candidate ranking results, wherein each group of topN candidate ranking results in the M+1 groups represents the N target voiceprints with the highest matching degree between the voiceprint model and the voiceprint library and the corresponding matching degree;

[0010] The topN candidate ranking results corresponding to each voiceprint model are updated based on the same target voiceprint and the corresponding matching degree in each group of topN candidate ranking results, and the updated topN candidate ranking results of each voiceprint model are output.

[0011] In a second aspect, an embodiment of the present application further provides a voiceprint retrieval device, the device comprising:

[0012] A segmentation module, used to segment the sound to be recognized by a preset segmentation strategy to obtain M sound segments to be recognized, where M is a positive integer greater than or equal to 2;

[0013] A construction module, used to extract features from the sound to be identified and the M sound segments to be identified, and construct M+1 voiceprint models based on the extraction results of the feature extraction;

[0014] A retrieval module is used to perform voiceprint retrieval on each of the M+1 voiceprint models to obtain M+1 groups of topN candidate ranking results, wherein each group of topN candidate ranking results in the M+1 groups represents the N target voiceprints with the highest matching degree between the voiceprint model and the voiceprint library and the corresponding matching degree;

[0015] An updating module is used to update the topN candidate ranking results corresponding to each voiceprint model based on the same target voiceprint and the corresponding matching degree in each group of topN candidate ranking results, and output the updated topN candidate ranking results of each voiceprint model.

[0016] In a third aspect, an embodiment of the present application further provides an electronic device, comprising: a processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium through the bus, and the processor executes the machine-readable instructions to execute the voiceprint retrieval method described in any one of the first aspects.

[0017] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the voiceprint retrieval method described in any one of the first aspects is executed.

[0018] The embodiments of the present application have the following beneficial effects:

[0019] (1) Improving the accuracy of voiceprint retrieval: By analyzing the length of the speech to be recognized and performing fine segmentation processing accordingly, each segment is independently subjected to feature extraction, voiceprint modeling, and voiceprint comparison. This method effectively avoids the performance degradation problem that may be caused by uneven distribution of voiceprint features or missing local features when processing the entire speech segment. The generation of multiple groups of topN candidate ranking results, combined with comprehensive consideration of the voiceprint features of different segments, makes the final ranking results more accurate and can more accurately screen out suspected hit candidate results that are highly matched with the speech information to be recognized.

[0020] (2) Enhance the robustness of voiceprint retrieval: Different segments may contain different voiceprint features. Some segments may more clearly reflect the specific characteristics of the speaker, while others may be less affected by noise or other interference. By combining the results of multiple groups of segments, the errors or deviations that may exist in a single segment can be eliminated, thereby enhancing the robustness of the voiceprint retrieval system and enabling it to maintain a high recognition accuracy in different environments.

[0021] (3) Improve retrieval efficiency and speed: Although segment processing increases the number of processing steps, each segment can be processed in parallel, and the feature extraction and voiceprint comparison of each segment are relatively independent, which helps to improve the overall retrieval efficiency. At the same time, through the complementary calculation of multiple groups of results, the most matching candidate results can be found in a shorter time, thereby speeding up the retrieval speed.

[0022] (4) Supporting applications in complex scenarios: In practical applications, the speech to be recognized often contains a variety of complex factors, such as background noise, speech speed changes, emotional fluctuations, etc. This method can better adapt to these complex scenarios through segmented processing and comprehensive calculation, and improve the practicality and reliability of voiceprint retrieval.

[0023] (5) Improve user experience: Due to the improvement in voiceprint retrieval accuracy, users can obtain accurate recognition results more quickly, thereby reducing waiting time and the trouble caused by misidentification.

[0024] In summary, this method effectively solves the performance degradation problem that may be caused by whole-segment speech processing in traditional voiceprint retrieval through strategies such as segmented processing and comprehensive calculation of multiple groups of results. It significantly improves the accuracy and robustness of voiceprint retrieval, while improving retrieval efficiency and speed, providing users with a better experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0026] Figure 1 It is a flowchart of steps S101-S104 provided in an embodiment of the present application;

[0027] Figure 2 It is a flowchart of steps S201-S204 provided in an embodiment of the present application;

[0028] Figure 3 It is a flowchart of steps S301-S303 provided in an embodiment of the present application;

[0029] Figure 4 It is a flowchart of steps S401-S402 provided in an embodiment of the present application;

[0030] Figure 5 It is a flowchart of steps S501-S503 provided in an embodiment of the present application;

[0031] Figure 6 is a schematic diagram of the structure of a voiceprint retrieval device provided in an embodiment of the present application;

[0032] Figure 7 It is a schematic diagram of the composition structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0033] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of explanation and description and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in this application shows the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowchart can be implemented out of sequence, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art can add one or more other operations to the flowchart under the guidance of the content of the present application, or remove one or more operations from the flowchart.

[0034] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0035] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0036] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0037] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the existence of the features declared thereafter, but does not exclude the addition of other features.

[0038] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0039] See also Figure 1 , Figure 1 is a flow chart of steps S101-S104 of the voiceprint retrieval method provided in the embodiment of the present application, which will be combined with Figure 1 Steps S101-S104 are shown for explanation.

[0040] In step S101, the sound to be recognized is segmented using a preset segmentation strategy to obtain M sound segments to be recognized, where M is a positive integer greater than or equal to 2.

[0041] First, the sound signal to be recognized is segmented by a preset segmentation strategy to obtain M sound segments to be recognized. The segmentation strategy here can be set according to the characteristics of the sound or a fixed time length. Segmentation processing helps to capture local features in the sound signal, especially when the sound signal is long or contains multiple different parts, segmentation processing can analyze each part more carefully, thereby improving the accuracy of recognition.

[0042] In step S102, feature extraction is performed on the sound to be recognized and the M sound segments to be recognized, and M+1 voiceprint models are constructed based on the extraction results of the feature extraction.

[0043] The original sound to be recognized and the M sound segments after segmentation are respectively subjected to feature extraction. These features may include spectral features, cepstral coefficients, Mel-frequency cepstral coefficients (MFCC), etc. Based on these features, M+1 voiceprint models are constructed (including the voiceprint models of the original sound and the M segments). Feature extraction is a key link in voiceprint recognition, which can convert sound signals into mathematical representations that can be used for comparison and recognition. Constructing multiple voiceprint models provides the possibility of analyzing sound signals from different angles and granularities, which helps to more comprehensively reflect the characteristics of the sound.

[0044] In step S103, voiceprint retrieval is performed on each of the M+1 voiceprint models to obtain M+1 groups of topN candidate ranking results, wherein each group of topN candidate ranking results in the M+1 groups represents the N target voiceprints with the highest matching degree between the voiceprint model and the voiceprint library and the corresponding matching degree.

[0045] Perform voiceprint retrieval on each voiceprint model, that is, search the voiceprint library for the N target voiceprints with the highest matching degree with the model, and generate the corresponding topN candidate ranking results. These results include the matching target voiceprints and their matching degrees. Through voiceprint retrieval, you can quickly locate the voiceprint model with a high similarity to the voice to be identified, providing strong support for subsequent identification or verification. The topN candidate ranking results further narrow the candidate range and improve the efficiency and accuracy of identification.

[0046] In step S104, the topN candidate ranking results corresponding to each voiceprint model are updated based on the same target voiceprint and the corresponding matching degree in each group of topN candidate ranking results, and the updated topN candidate ranking results of each voiceprint model are output.

[0047] Based on the same target voiceprint and its matching degree in each group of topN candidate ranking results, the topN candidate ranking results corresponding to each voiceprint model are updated. The updated results are ranked in order based on the average value of the duplicate items. Finally, the updated topN candidate ranking results of each voiceprint model are output. By updating the results, the information from different voiceprint models can be further integrated, redundancy and errors can be reduced, and the reliability and consistency of the final recognition results can be improved.

[0048] In some embodiments, the preset segmentation strategy is determined based on at least one of segment length, the number of sounds in the sound to be recognized, and valid speech in the sound to be recognized.

[0049] In some embodiments, the sound to be recognized is segmented using a preset segmentation strategy, including:

[0050] When segmentation processing is performed based on the segment length, determining the maximum length of the sound segment to be recognized, and segmenting the sound to be recognized according to the maximum length;

[0051] In this method, a maximum length is first determined (for example, a fixed number of seconds or a certain number of sample points), and then the sound to be recognized is divided into equal parts according to this maximum length. If the total length of the sound to be recognized is not an integer multiple of the maximum length, the last segment will be shorter than the maximum length.

[0052] When segmentation processing is performed based on the number of sounds in the to-be-recognized sound, voice data of the same type / person are separated by a voice separation model;

[0053] Here, when the sound to be recognized contains multiple speakers or different types of sounds, the speech separation model can be used to separate these sounds first. According to the parameter setting value, it is determined whether to use the speech separation technology to separate the original multiple speakers' recorded speech data to form a single speaker's speech group. (1) The default number of speakers is 1, which assumes that the speech to be recognized contains only one speaker, and there is no need to perform multi-speaker speech separation, and directly perform segmentation processing; (2) When the number of speakers is set to X, it is assumed that the audio to be recognized contains X speakers, and the original recorded speech data is first separated by speech separation technology to form X speakers' speech groups, and each speech group is then segmented.

[0054] When segmentation processing is performed based on the effective speech in the sound to be recognized, the effective speech length is determined, the effective speech in the sound to be recognized is extracted, and the effective speech is segmented according to the effective speech length.

[0055] In this method, the effective speech area in the sound to be recognized can be determined by VAD technology, and then only these effective speech areas are segmented. The sound to be recognized is detected by VAD technology to obtain the effective speech area. The total length Lvalid of the effective speech is determined. The appropriate segmentation strategy is selected according to the needs to segment the effective speech.

[0056] It should be noted that these segmentation methods can be used in combination with each other to more flexibly deal with complex and changing sound signals. Combining different segmentation strategies can make full use of their respective advantages and make up for their respective shortcomings, thereby improving the accuracy and efficiency of voiceprint retrieval.

[0057] In some embodiments, see Figure 2 , Figure 2 It is a flowchart of steps S201-S204 provided in an embodiment of the present application. The feature extraction of the sound to be identified and the M sound segments to be identified can be achieved through steps S201-S204, which will be explained in combination with each step.

[0058] In step S201, the speech spectrum information of each of the M sound segments to be recognized is obtained by at least one of fast Fourier transform FFT, discrete cosine transform DCT, and wavelet transform.

[0059] In step S202, valid information is extracted from the speech spectrum information of each sound segment to be recognized.

[0060] In step S203, denoising is performed on the effective information.

[0061] In step S204, feature parameters for characterizing the voiceprint characteristics are extracted based on the effective information after denoising.

[0062] Here, each of the M sound segments to be recognized is first transformed by at least one of Fast Fourier Transform (FFT), Discrete Cosine Transform (DCT) or Wavelet Transform to obtain its speech spectrum information.

[0063] Fast Fourier Transform (FFT): It is one of the most commonly used spectrum analysis methods, which can convert sound signals from time domain to frequency domain, so as to observe the frequency distribution of sound signals.

[0064] Discrete Cosine Transform (DCT): Although not directly used for spectral analysis, it is widely used in sound coding and compression, and sometimes for feature extraction because it can effectively represent the energy distribution in the signal.

[0065] Wavelet transform: It provides a multi-scale analysis method that can simultaneously provide information about the signal in both the time domain and the frequency domain, and is particularly useful for the analysis of non-stationary signals.

[0066] Next, from the speech spectrum information of each sound segment to be recognized, information useful for voiceprint recognition needs to be extracted. This usually includes the envelope of the spectrum, energy distribution, harmonic structure, etc. The specific information to be extracted depends on the requirements of the subsequent voiceprint model and the design of the recognition algorithm.

[0067] Since actual sound signals often contain various noises (such as environmental noise, equipment noise, etc.), these noises will interfere with the extraction and recognition of voiceprint features. Therefore, in the denoising process, appropriate methods need to be used to remove or reduce the impact of noise. Common denoising methods include spectral subtraction, Wiener filtering, wavelet denoising, etc.

[0068] Based on the effective information after denoising, feature parameters are extracted to characterize the voiceprint characteristics. These feature parameters should be able to stably reflect the speaker's personality characteristics and have sufficient discrimination between different speakers. Commonly used voiceprint feature parameters include Mel Frequency Cepstral Coefficients (MFCC), Linear Prediction Cepstral Coefficients (LPCC), Pitch Contour, etc.

[0069] Mel-frequency cepstral coefficient (MFCC): It is one of the most commonly used features in voiceprint recognition. It simulates the auditory characteristics of the human ear and performs nonlinear transformation and cepstral analysis on the sound signal.

[0070] Linear Prediction Cepstral Coefficients (LPCC): Based on the Linear Prediction Coding (LPC) technology, it estimates future samples by predicting the current samples of the sound signal, thereby extracting the dynamic characteristics of the sound signal.

[0071] Pitch Contour: It reflects the changes in the fundamental frequency in the sound signal and is of great significance for distinguishing the voice characteristics of different speakers.

[0072] In some embodiments, see Figure 3 , Figure 3 It is a flowchart of steps S301-S303 provided in an embodiment of the present application. The construction of M+1 voiceprint models based on the extraction results of the feature extraction can be achieved through steps S301-S303, which will be explained in combination with each step.

[0073] In step S301, cluster analysis is performed on the characteristic parameters to obtain a distribution space for describing the characteristic parameters.

[0074] In step S302, the optimal subspace expression of the sound segment to be recognized is learned based on the i-vector subspace learning model to obtain an i-vector expression matrix, and the i-vector vector features are calculated through the i-vector expression matrix.

[0075] In step S303, M+1 voiceprint models are obtained based on i-vector vector feature training through linear discriminant analysis or probabilistic linear discriminant analysis.

[0076] Here, the extracted feature parameters are first subjected to cluster analysis. Cluster analysis is an unsupervised learning method that is used to divide samples in a data set into several groups or clusters, so that the samples in the same cluster have a high similarity, while the samples in different clusters have a low similarity. In voiceprint recognition, cluster analysis can understand the distribution space of feature parameters and provide a basis for subsequent subspace learning and model training. Next, the optimal subspace expression of the sound segment to be identified is learned based on the i-vector subspace learning model. i-vector is a low-dimensional vector representation method for voiceprint recognition. It captures the personality characteristics of the speaker by mapping the feature parameters of the sound signal into a low-dimensional subspace. In the i-vector subspace learning process, joint factor analysis (JFA) or similar statistical models are usually used to estimate the speaker space and channel space, and calculate the i-vector expression of each sound segment. This i-vector expression matrix contains the coordinate information of the sound segment in the i-vector subspace, which can be used for subsequent voiceprint model training or recognition. Finally, M+1 voiceprint models are trained based on i-vector features through methods such as linear discriminant analysis (LDA) or probabilistic linear discriminant analysis (PLDA). LDA is a supervised learning method used to find a linear transformation so that the transformed samples have the largest inter-class dispersion and the smallest intra-class dispersion. In voiceprint recognition, LDA can be used to map i-vector features into a more discriminative space, thereby constructing a voiceprint model that can characterize the characteristics of different speakers. PLDA is an extension of LDA that takes into account the prior distribution of intra-class and inter-class covariance matrices and introduces a probabilistic model to deal with uncertainty. In voiceprint recognition, PLDA is usually used in the scoring stage to calculate the similarity or distance between two voiceprint models.

[0077] In some embodiments, see Figure 4 , Figure 4 It is a flow chart of steps S401-S402 provided in an embodiment of the present application, wherein voiceprint retrieval is performed on each of the M+1 voiceprint models respectively to obtain M+1 groups of topN candidate ranking results, which can be achieved through steps S401-S402 and will be explained in combination with each step.

[0078] In step S401, each voiceprint model is input into a preset voiceprint library.

[0079] In step S402, for each voiceprint model, the most similar N target voiceprints are retrieved from the voiceprint library as candidates, and are ranked from high to low according to the matching degree.

[0080] Here, first ensure that there is a preset voiceprint library, which contains voiceprint models of a large number of known speakers. Input M+1 voiceprint models into the preset voiceprint library respectively. In actual operation, the parameters of each voiceprint model (such as i-vector vector, GMM parameters, etc.) are usually input into the retrieval system as a query. For each input voiceprint model, the retrieval system searches for the most similar N target voiceprints in the voiceprint library. The measure of similarity is usually based on some distance or similarity function, such as cosine similarity, Euclidean distance, Mahalanobis distance, etc., which are calculated based on the characteristic parameters of the voiceprint model. According to the calculated similarity or distance, the retrieved N target voiceprints are ranked from high to low according to the matching degree. The voiceprint with the highest matching degree is ranked first, and so on. For each input voiceprint model, a set of ranking results containing N candidate voiceprints will be obtained. In this way, M+1 groups of topN candidate ranking results are obtained.

[0081] In some embodiments, see Figure 5 , Figure 5 It is a flow chart of steps S501-S503 provided in an embodiment of the present application. The topN candidate ranking results corresponding to each voiceprint model are updated based on the same target voiceprint in each group of topN candidate ranking results and the corresponding matching degree. This can be achieved through steps S501-S503, which will be explained in combination with each step.

[0082] In step S501, an average value of the matching degree of the same target voiceprint is calculated based on the same target voiceprint in each group of topN candidate ranking results.

[0083] In step S502, the matching degree of the same target voiceprint is updated based on the average value.

[0084] In step S503, the topN candidate ranking results of each voiceprint model are reordered based on the updated matching degree to obtain the updated topN candidate ranking results of each voiceprint model.

[0085] Here, after obtaining the topN candidate ranking results for each voiceprint model, in order to further improve the accuracy and robustness of recognition, the common information in multiple groups of results can be used to update these ranking results. In particular, when multiple voiceprint models contain the same target voiceprint in the topN candidates, we can update their matching degree by calculating the average matching degree of these common target voiceprints and re-rank the candidate list accordingly.

[0086] Specifically, calculate the average matching degree of the same target voiceprint: traverse the topN candidate ranking results of all voiceprint models. For each target voiceprint that appears at least twice, calculate the average matching degree of all relevant voiceprint models. Use the calculated average as the updated matching degree to replace the original matching degree. Reorder the topN candidate ranking results of each voiceprint model according to the updated matching degree.

[0087] In summary, the embodiments of the present application have the following beneficial effects:

[0088] (1) Improving the accuracy of voiceprint retrieval: By analyzing the length of the speech to be recognized and performing fine segmentation processing accordingly, each segment is independently subjected to feature extraction, voiceprint modeling, and voiceprint comparison. This method effectively avoids the performance degradation problem that may be caused by uneven distribution of voiceprint features or missing local features when processing the entire speech segment. The generation of multiple groups of topN candidate ranking results, combined with comprehensive consideration of the voiceprint features of different segments, makes the final ranking results more accurate and can more accurately screen out suspected hit candidate results that are highly matched with the speech information to be recognized.

[0089] (2) Enhance the robustness of voiceprint retrieval: Different segments may contain different voiceprint features. Some segments may more clearly reflect the specific characteristics of the speaker, while others may be less affected by noise or other interference. By combining the results of multiple groups of segments, the errors or deviations that may exist in a single segment can be eliminated, thereby enhancing the robustness of the voiceprint retrieval system and enabling it to maintain a high recognition accuracy in different environments.

[0090] (3) Improve retrieval efficiency and speed: Although segment processing increases the number of processing steps, each segment can be processed in parallel, and the feature extraction and voiceprint comparison of each segment are relatively independent, which helps to improve the overall retrieval efficiency. At the same time, through the complementary calculation of multiple groups of results, the most matching candidate results can be found in a shorter time, thereby speeding up the retrieval speed.

[0091] (4) Supporting applications in complex scenarios: In practical applications, the speech to be recognized often contains a variety of complex factors, such as background noise, speech speed changes, emotional fluctuations, etc. This method can better adapt to these complex scenarios through segmented processing and comprehensive calculation, and improve the practicality and reliability of voiceprint retrieval.

[0092] (5) Improve user experience: Due to the improvement in voiceprint retrieval accuracy, users can obtain accurate recognition results more quickly, thereby reducing waiting time and the trouble caused by misidentification.

[0093] In summary, this method effectively solves the performance degradation problem that may be caused by whole-segment speech processing in traditional voiceprint retrieval through strategies such as segmented processing and comprehensive calculation of multiple groups of results. It significantly improves the accuracy and robustness of voiceprint retrieval, while improving retrieval efficiency and speed, providing users with a better experience.

[0094] Based on the same inventive concept, the embodiments of the present application also provide a voiceprint retrieval device corresponding to the voiceprint retrieval method in the first embodiment. Since the principle of solving the problem by the device in the embodiments of the present application is similar to that of the above-mentioned voiceprint retrieval method, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0095] like Figure 6 As shown, Figure 6 : is a schematic diagram of the structure of a voiceprint retrieval device 600 provided in an embodiment of the present application. The voiceprint retrieval device 600 includes:

[0096] The segmentation module 601 is used to segment the sound to be recognized by using a preset segmentation strategy to obtain M sound segments to be recognized, where M is a positive integer greater than or equal to 2;

[0097] A construction module 602 is used to extract features from the sound to be recognized and the M sound segments to be recognized, and to construct M+1 voiceprint models based on the extraction results of the feature extraction;

[0098] The retrieval module 603 is used to perform voiceprint retrieval on each of the M+1 voiceprint models to obtain M+1 groups of topN candidate ranking results, wherein each group of topN candidate ranking results in the M+1 groups represents the N target voiceprints with the highest matching degree between the voiceprint model and the voiceprint library and the corresponding matching degree;

[0099] The updating module 604 is used to update the topN candidate ranking result corresponding to each voiceprint model based on the same target voiceprint and the corresponding matching degree in each group of topN candidate ranking results, and output the updated topN candidate ranking result of each voiceprint model.

[0100] Those skilled in the art should understand that Figure 6 The implementation functions of each unit in the voiceprint retrieval device 600 shown can be understood by referring to the relevant description of the aforementioned voiceprint retrieval method. Figure 6 The functions of each unit in the voiceprint retrieval device 600 shown can be implemented by a program running on a processor, or by a specific logic circuit.

[0101] In a possible implementation, the preset segmentation strategy is determined based on at least one of segment length, the number of sounds in the sound to be recognized, and valid speech in the sound to be recognized.

[0102] In a possible implementation, the segmentation module 601 performs segmentation processing on the sound to be recognized according to a preset segmentation strategy, including:

[0103] When segmentation processing is performed based on the segment length, determining the maximum length of the sound segment to be recognized, and segmenting the sound to be recognized according to the maximum length;

[0104] When segmentation processing is performed based on the number of sounds in the to-be-recognized sound, voice data of the same type / person are separated by a voice separation model;

[0105] When segmentation processing is performed based on the effective speech in the sound to be recognized, the effective speech length is determined, the effective speech in the sound to be recognized is extracted, and the effective speech is segmented according to the effective speech length.

[0106] In a possible implementation, the construction module 602 performs feature extraction on the sound to be recognized and the M sound segments to be recognized, respectively, including:

[0107] Acquire speech spectrum information of each of the M sound segments to be recognized by at least one of fast Fourier transform FFT, discrete cosine transform DCT, and wavelet transform;

[0108] Extracting effective information from the speech spectrum information of each sound segment to be recognized;

[0109] Performing denoising processing on the effective information;

[0110] Feature parameters for characterizing the voiceprint characteristics are extracted based on the effective information after denoising.

[0111] In a possible implementation, the construction module 602 constructs M+1 voiceprint models based on the extraction result of the feature extraction, including:

[0112] Performing cluster analysis on the characteristic parameters to obtain a distribution space for describing the characteristic parameters;

[0113] Learning the optimal subspace expression of the sound segment to be recognized based on the i-vector subspace learning model to obtain an i-vector expression matrix, and calculating the i-vector vector features through the i-vector expression matrix;

[0114] Through linear discriminant analysis or probabilistic linear discriminant analysis, M+1 voiceprint models are obtained based on i-vector vector feature training.

[0115] In a possible implementation, the retrieval module 603 performs voiceprint retrieval on each of the M+1 voiceprint models respectively to obtain M+1 groups of topN candidate ranking results, including:

[0116] Input each voiceprint model into a preset voiceprint library;

[0117] For each voiceprint model, the most similar N target voiceprints are retrieved from the voiceprint library as candidates, and are ranked from high to low according to the matching degree.

[0118] In a possible implementation, the updating module 604 updates the topN candidate ranking results corresponding to each voiceprint model based on the same target voiceprint and the corresponding matching degree in each group of topN candidate ranking results, including:

[0119] Calculate the average matching degree of the same target voiceprint based on the same target voiceprint in each group of topN candidate ranking results;

[0120] Updating the matching degree of the same target voiceprint based on the average value;

[0121] The topN candidate ranking results of each voiceprint model are reordered based on the updated matching degree to obtain the updated topN candidate ranking results of each voiceprint model.

[0122] The above-mentioned voiceprint retrieval device has the following beneficial effects:

[0123] (1) Improving the accuracy of voiceprint retrieval: By analyzing the length of the speech to be recognized and performing fine segmentation processing accordingly, each segment is independently subjected to feature extraction, voiceprint modeling, and voiceprint comparison. This method effectively avoids the performance degradation problem that may be caused by uneven distribution of voiceprint features or missing local features when processing the entire speech segment. The generation of multiple groups of topN candidate ranking results, combined with comprehensive consideration of the voiceprint features of different segments, makes the final ranking results more accurate and can more accurately screen out suspected hit candidate results that are highly matched with the speech information to be recognized.

[0124] (2) Enhance the robustness of voiceprint retrieval: Different segments may contain different voiceprint features. Some segments may more clearly reflect the specific characteristics of the speaker, while others may be less affected by noise or other interference. By combining the results of multiple groups of segments, the errors or deviations that may exist in a single segment can be eliminated, thereby enhancing the robustness of the voiceprint retrieval system and enabling it to maintain a high recognition accuracy in different environments.

[0125] (3) Improve retrieval efficiency and speed: Although segment processing increases the number of processing steps, each segment can be processed in parallel, and the feature extraction and voiceprint comparison of each segment are relatively independent, which helps to improve the overall retrieval efficiency. At the same time, through the complementary calculation of multiple groups of results, the most matching candidate results can be found in a shorter time, thereby speeding up the retrieval speed.

[0126] (4) Supporting applications in complex scenarios: In practical applications, the speech to be recognized often contains a variety of complex factors, such as background noise, speech speed changes, emotional fluctuations, etc. This method can better adapt to these complex scenarios through segmented processing and comprehensive calculation, and improve the practicality and reliability of voiceprint retrieval.

[0127] (5) Improve user experience: Due to the improvement in voiceprint retrieval accuracy, users can obtain accurate recognition results more quickly, thereby reducing waiting time and the trouble caused by misidentification.

[0128] In summary, this method effectively solves the performance degradation problem that may be caused by whole-segment speech processing in traditional voiceprint retrieval through strategies such as segmented processing and comprehensive calculation of multiple groups of results. It significantly improves the accuracy and robustness of voiceprint retrieval, while improving retrieval efficiency and speed, providing users with a better experience.

[0129] like Figure 7 As shown, Figure 7 The present invention provides a schematic diagram of the structure of an electronic device 700, wherein the electronic device 700 includes:

[0130] A processor 701, a storage medium 702 and a bus 703, wherein the storage medium 702 stores machine-readable instructions executable by the processor 701. When the electronic device 700 is running, the processor 701 communicates with the storage medium 702 via the bus 703, and the processor 701 executes the machine-readable instructions to perform the steps of the voiceprint retrieval method described in the embodiment of the present application.

[0131] In actual application, the components in the electronic device 700 are coupled together via the bus 703. It is understood that the bus 703 is used to realize the connection and communication between these components. In addition to the data bus, the bus 703 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, Figure 7 Various buses are labeled as bus 703.

[0132] The electronic device has the following beneficial effects:

[0133] (1) Improving the accuracy of voiceprint retrieval: By analyzing the length of the speech to be recognized and performing fine segmentation processing accordingly, each segment is independently subjected to feature extraction, voiceprint modeling, and voiceprint comparison. This method effectively avoids the performance degradation problem that may be caused by uneven distribution of voiceprint features or missing local features when processing the entire speech segment. The generation of multiple groups of topN candidate ranking results, combined with comprehensive consideration of the voiceprint features of different segments, makes the final ranking results more accurate and can more accurately screen out suspected hit candidate results that are highly matched with the speech information to be recognized.

[0134] (2) Enhance the robustness of voiceprint retrieval: Different segments may contain different voiceprint features. Some segments may more clearly reflect the specific characteristics of the speaker, while others may be less affected by noise or other interference. By combining the results of multiple groups of segments, the errors or deviations that may exist in a single segment can be eliminated, thereby enhancing the robustness of the voiceprint retrieval system and enabling it to maintain a high recognition accuracy in different environments.

[0135] (3) Improve retrieval efficiency and speed: Although segment processing increases the number of processing steps, each segment can be processed in parallel, and the feature extraction and voiceprint comparison of each segment are relatively independent, which helps to improve the overall retrieval efficiency. At the same time, through the complementary calculation of multiple groups of results, the most matching candidate results can be found in a shorter time, thereby speeding up the retrieval speed.

[0136] (4) Supporting applications in complex scenarios: In practical applications, the speech to be recognized often contains a variety of complex factors, such as background noise, speech speed changes, emotional fluctuations, etc. This method can better adapt to these complex scenarios through segmented processing and comprehensive calculation, and improve the practicality and reliability of voiceprint retrieval.

[0137] (5) Improve user experience: Due to the improvement in voiceprint retrieval accuracy, users can obtain accurate recognition results more quickly, thereby reducing waiting time and the trouble caused by misidentification.

[0138] In summary, this method effectively solves the performance degradation problem that may be caused by whole-segment speech processing in traditional voiceprint retrieval through strategies such as segmented processing and comprehensive calculation of multiple groups of results. It significantly improves the accuracy and robustness of voiceprint retrieval, while improving retrieval efficiency and speed, providing users with a better experience.

[0139] The embodiment of the present application further provides a computer-readable storage medium, which stores executable instructions. When the executable instructions are executed by at least one processor 701, the voiceprint retrieval method described in the embodiment of the present application is implemented.

[0140] In some embodiments, the storage medium can be a magnetic random access memory (FRAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory, a magnetic surface storage, an optical disk, or a compact disc read-only memory (CD-ROM); it can also be various devices including one or any combination of the above memories.

[0141] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0142] As an example, executable instructions may, but do not necessarily, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (for example, files storing one or more modules, subroutines, or code portions).

[0143] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0144] The computer-readable storage medium has the following advantages:

[0145] (1) Improving the accuracy of voiceprint retrieval: By analyzing the length of the speech to be recognized and performing fine segmentation processing accordingly, each segment is independently subjected to feature extraction, voiceprint modeling, and voiceprint comparison. This method effectively avoids the performance degradation problem that may be caused by uneven distribution of voiceprint features or missing local features when processing the entire speech segment. The generation of multiple groups of topN candidate ranking results, combined with comprehensive consideration of the voiceprint features of different segments, makes the final ranking results more accurate and can more accurately screen out suspected hit candidate results that are highly matched with the speech information to be recognized.

[0146] (2) Enhance the robustness of voiceprint retrieval: Different segments may contain different voiceprint features. Some segments may more clearly reflect the specific characteristics of the speaker, while others may be less affected by noise or other interference. By combining the results of multiple groups of segments, the errors or deviations that may exist in a single segment can be eliminated, thereby enhancing the robustness of the voiceprint retrieval system and enabling it to maintain a high recognition accuracy in different environments.

[0147] (3) Improve retrieval efficiency and speed: Although segment processing increases the number of processing steps, each segment can be processed in parallel, and the feature extraction and voiceprint comparison of each segment are relatively independent, which helps to improve the overall retrieval efficiency. At the same time, through the complementary calculation of multiple groups of results, the most matching candidate results can be found in a shorter time, thereby speeding up the retrieval speed.

[0148] (4) Supporting applications in complex scenarios: In practical applications, the speech to be recognized often contains a variety of complex factors, such as background noise, speech speed changes, emotional fluctuations, etc. This method can better adapt to these complex scenarios through segmented processing and comprehensive calculation, and improve the practicality and reliability of voiceprint retrieval.

[0149] (5) Improve user experience: Due to the improvement in voiceprint retrieval accuracy, users can obtain accurate recognition results more quickly, thereby reducing waiting time and the trouble caused by misidentification.

[0150] In summary, this method effectively solves the performance degradation problem that may be caused by whole-segment speech processing in traditional voiceprint retrieval through strategies such as segmented processing and comprehensive calculation of multiple groups of results. It significantly improves the accuracy and robustness of voiceprint retrieval, while improving retrieval efficiency and speed, providing users with a better experience.

[0151] In the several embodiments provided in the present application, it should be understood that the disclosed methods and electronic devices can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0152] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0153] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0154] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, 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, a platform server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard drives, ROM, RAM, magnetic disks, or optical disks.

[0155] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A voiceprint retrieval method, characterized in that: The method comprises: Segment the sound to be recognized by using a preset segmentation strategy to obtain M sound segments to be recognized, where M is a positive integer greater than or equal to 2; Performing feature extraction on the sound to be identified and the M sound segments to be identified respectively, and constructing M+1 voiceprint models based on the extraction results of the feature extraction; Perform voiceprint retrieval on each of the M+1 voiceprint models respectively to obtain M+1 groups of topN candidate ranking results, wherein each group of topN candidate ranking results in the M+1 groups represents the N target voiceprints with the highest matching degree between the voiceprint model and the voiceprint library and the corresponding matching degree; Based on the same target voiceprint and the corresponding matching degree in each group of topN candidate ranking results, the topN candidate ranking results corresponding to each voiceprint model are updated, and the updated topN candidate ranking results of each voiceprint model are output; The preset segmentation strategy is determined based on at least one of the segment length, the number of sounds in the sound to be recognized, and the valid speech in the sound to be recognized; The sound to be recognized is segmented using a preset segmentation strategy, including: When segmentation processing is performed based on the segment length, determining the maximum length of the sound segment to be recognized, and segmenting the sound to be recognized according to the maximum length; When segmentation processing is performed based on the number of sounds in the to-be-recognized sound, voice data of the same type / person are separated by a voice separation model; When segmenting the effective speech in the sound to be recognized, determining the effective speech length, extracting the effective speech in the sound to be recognized, and segmenting the effective speech according to the effective speech length; The extracting features of the sound to be recognized and the M sound segments to be recognized respectively includes: Acquire speech spectrum information of each of the M sound segments to be recognized by at least one of fast Fourier transform FFT, discrete cosine transform DCT, and wavelet transform; Extracting effective information from the speech spectrum information of each sound segment to be recognized; Performing denoising processing on the effective information; Extracting characteristic parameters for characterizing the voiceprint characteristics based on the effective information after denoising; The constructing M+1 voiceprint models based on the extraction results of the feature extraction includes: Performing cluster analysis on the characteristic parameters to obtain a distribution space for describing the characteristic parameters; Learning the optimal subspace expression of the sound segment to be recognized based on the i-vector subspace learning model to obtain an i-vector expression matrix, and calculating the i-vector vector features through the i-vector expression matrix; Through linear discriminant analysis or probabilistic linear discriminant analysis, M+1 voiceprint models are obtained based on i-vector vector feature training; The voiceprint retrieval is performed on each of the M+1 voiceprint models to obtain M+1 groups of topN candidate ranking results, including: Input each voiceprint model into a preset voiceprint library; For each voiceprint model, retrieve the most similar N target voiceprints from the voiceprint library as candidates, and rank them from high to low according to the matching degree; The updating of the topN candidate ranking results corresponding to each voiceprint model based on the same target voiceprint and the corresponding matching degree in each group of topN candidate ranking results includes: Calculate the average matching degree of the same target voiceprint based on the same target voiceprint in each group of topN candidate ranking results; Updating the matching degree of the same target voiceprint based on the average value; The topN candidate ranking results of each voiceprint model are reordered based on the updated matching degree to obtain the updated topN candidate ranking results of each voiceprint model.

2. A voiceprint retrieval device, characterized in that: The device comprises: A segmentation module is used to segment the sound to be recognized by a preset segmentation strategy to obtain M sound segments to be recognized, wherein M is a positive integer greater than or equal to 2; the preset segmentation strategy is determined based on at least one of the segmentation length, the number of sounds in the sound to be recognized, and the valid speech in the sound to be recognized; the sound to be recognized is segmented by the preset segmentation strategy, including: when segmenting based on the segmentation length, determining the maximum length of the sound segment to be recognized, and segmenting the sound to be recognized according to the maximum length; when segmenting based on the number of sounds in the sound to be recognized, separating the speech data of the same type / person through a speech separation model; when segmenting based on the valid speech in the sound to be recognized, determining the effective speech length, extracting the valid speech in the sound to be recognized, and segmenting the valid speech according to the effective speech length; A construction module is used to perform feature extraction on the sound to be identified and the M sound segments to be identified, and construct M+1 voiceprint models based on the extraction results of the feature extraction; the feature extraction on the sound to be identified and the M sound segments to be identified, respectively, includes: obtaining the speech spectrum information of each of the M sound segments to be identified through at least one of fast Fourier transform FFT, discrete cosine DCT transform, and wavelet transform; extracting effective information from the speech spectrum information of each sound segment to be identified; denoising the effective information; based on the denoised Effective information is used to extract characteristic parameters for characterizing voiceprint characteristics; the extraction result based on the characteristic extraction constructs M+1 voiceprint models, including: performing cluster analysis on the characteristic parameters to obtain a distribution space for describing the characteristic parameters; learning the optimal subspace expression of the sound segment to be identified based on the i-vector subspace learning model to obtain an i-vector expression matrix, and calculating the i-vector vector features through the i-vector expression matrix; obtaining M+1 voiceprint models based on i-vector vector feature training through linear discriminant analysis or probabilistic linear discriminant analysis; A retrieval module is used to perform voiceprint retrieval on each of the M+1 voiceprint models to obtain M+1 groups of topN candidate ranking results, wherein each group of topN candidate ranking results in the M+1 groups represents the N target voiceprints with the highest matching degree between the voiceprint model and the voiceprint library and the corresponding matching degree; performing voiceprint retrieval on each of the M+1 voiceprint models to obtain M+1 groups of topN candidate ranking results comprises: inputting each of the voiceprint models into a preset voiceprint library; for each of the voiceprint models, retrieving the most similar N target voiceprints from the voiceprint library as candidates, and ranking them from high to low according to the matching degree; An updating module is used to update the topN candidate ranking results corresponding to each voiceprint model based on the same target voiceprint and the corresponding matching degree in each group of topN candidate ranking results, and output the updated topN candidate ranking results of each voiceprint model; the updating of the topN candidate ranking results corresponding to each voiceprint model based on the same target voiceprint and the corresponding matching degree in each group of topN candidate ranking results includes: calculating the average value of the matching degree of the same target voiceprint based on the same target voiceprint in each group of topN candidate ranking results; updating the matching degree of the same target voiceprint based on the average value; reordering the topN candidate ranking results of each voiceprint model based on the updated matching degree to obtain the updated topN candidate ranking results of each voiceprint model.

3. An electronic device, characterized in that: include: A processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to execute the voiceprint retrieval method as described in claim 1.

4. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the voiceprint retrieval method according to claim 1 is executed.

Images