A speaker recognition method, apparatus, storage medium and device

Abstract

The application discloses a speaker recognition method and device, a storage medium and equipment. The method comprises the following steps: firstly, obtaining target speech to be recognized, and extracting acoustic features thereof; then, constructing an acoustic feature spectrogram corresponding to the target speech by using the acoustic features, and cutting the acoustic feature spectrogram into N spectrogram picture segments according to a preset window length; then, inputting the N spectrogram picture segments into a pre-constructed speaker recognition model to recognize a target voiceprint feature vector of a target speaker; and further, recognizing the target speaker according to the target voiceprint feature vector to obtain a recognition result of the target speaker. Since the speaker recognition model is constructed by using acoustic features in an original feature domain and a spectrum augmented feature domain and jointly training the acoustic features and cross-domain feature information, the voiceprint feature extraction effect of the model is effectively improved, so that the accuracy of the recognition result can be effectively improved when the speaker recognition model is used for speaker recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech processing, and particularly relates to a speaker recognition method and device, a storage medium and equipment. BACKGROUND

[0002] Speech interaction as an important human-computer interaction method can bring great convenience to people's life, and the more important one is the technology of identifying a speaker based on speech, which is called speaker recognition. For example, speaker recognition can be applied to occasions for confirming the identity of a speaker, such as court trials, remote financial services, security, etc., which all need to accurately identify the identity of a speaker based on speech data. Among them, voiceprint recognition is one of the key technologies in speaker identity recognition.

[0003] The existing method of identifying the identity of a speaker based on a voiceprint recognition result usually adopts a voiceprint modeling method based on a convolutional neural network (CNN), but the voiceprint recognition model established at present has the problem of mismatch between acoustic feature domains in the training and use stages, resulting in a low accuracy of the recognition result of the model.

[0004] Therefore, how to improve the accuracy of the voiceprint recognition result and further improve the accuracy of the speaker recognition result is a technical problem to be solved at present. SUMMARY

[0005] The main purpose of the embodiments of the present application is to provide a speaker recognition method and device, a storage medium and equipment, which can effectively improve the accuracy of the voiceprint recognition result when performing speaker recognition, and further improve the accuracy of the speaker recognition result.

[0006] The embodiments of the present application provide a speaker recognition method, which comprises:

[0007] Obtaining a target speech to be recognized, and extracting acoustic features of the target speech;

[0008] Using the acoustic features of the target speech, constructing an acoustic feature spectrogram corresponding to the target speech, and cutting the acoustic feature spectrogram into N spectrogram picture segments according to a preset window length; the N is a positive integer greater than 0;

[0009] Inputting the N spectrogram picture segments into a pre-constructed speaker recognition model to identify a target voiceprint feature vector of a target speaker; the speaker recognition model is obtained by jointly training acoustic features in an original feature domain and a spectrum augmented feature domain, and combining cross-domain feature information;

[0010] Based on the target voiceprint feature vector, the target speaker is identified to obtain the identification result of the target speaker.

[0011] In one possible implementation, the step of inputting the N spectrogram image segments into a pre-built speaker recognition model to identify the target speaker's target voiceprint feature vector includes:

[0012] The N spectrogram image segments are sequentially input into a pre-built speaker recognition model to obtain N voiceprint feature vectors;

[0013] The N voiceprint feature vectors are weighted and averaged, and the result is used as the target voiceprint feature vector of the target speaker.

[0014] In one possible implementation, the method further includes:

[0015] Using sample speech and a preset target loss function, the initial speaker recognition model is trained across domains to obtain the trained speaker recognition model.

[0016] The trained speaker recognition model is validated and updated using the validation speech to obtain the updated speaker recognition model.

[0017] In one possible implementation, the initial speaker recognition model is the mainstream voiceprint recognition model, ECAPA neural network.

[0018] In one possible implementation, the step of using sample speech and a preset target loss function to perform cross-domain training on the initial speaker recognition model to obtain the trained speaker recognition model includes:

[0019] The acoustic features of the sample speech are extracted, and the acoustic features of the sample speech are used to construct the sample acoustic feature spectrogram corresponding to the sample speech. The sample acoustic feature spectrogram is then divided into M sample spectrogram image segments according to a preset window length; where M is a positive integer greater than 0.

[0020] The M sample spectrogram image segments are subjected to online spectrogram augmentation to obtain M sample spectrogram image segments after spectrogram augmentation.

[0021] The M sample spectrogram image segments and the M sample spectrogram image segments after spectral augmentation are input into the initial speaker recognition model in pairs to obtain M sample voiceprint feature vectors and M spectral augmented sample voiceprint feature vectors, respectively.

[0022] Using the M sample voiceprint feature vectors and the M spectral augmented sample voiceprint feature vectors, construct triplet voiceprint feature vectors in the original acoustic feature domain and triplet voiceprint feature vectors in the spectral augmented feature domain; and construct triplet voiceprint feature vectors in different acoustic feature domains.

[0023] The initial speaker recognition model is trained based on the triplet voiceprint feature vectors in the original acoustic feature domain, the triplet voiceprint feature vectors in the augmented feature domain, the triplet voiceprint feature vectors in different acoustic feature domains, and the preset target loss function, to obtain the trained speaker recognition model.

[0024] In one possible implementation, the preset target loss function includes a first target loss function, a second target loss function, and a third target loss function; the first target loss function is used to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vector under the same acoustic feature domain, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector; the second target loss function is used to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vector under different acoustic feature domains, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector; the third target loss function is used to improve the accuracy of the speaker recognition model's recognition result.

[0025] In one possible implementation, the acoustic features of the target speech are Mel-frequency cepstral coefficients (MFCCs) or filter bank features.

[0026] This application also provides a speaker recognition device, including:

[0027] An acquisition unit is used to acquire the target speech to be identified and extract the acoustic features of the target speech;

[0028] The segmentation unit is used to construct an acoustic feature spectrogram corresponding to the target speech using the acoustic features of the target speech, and to segment the acoustic feature spectrogram into N spectrogram image segments according to a preset window length; where N is a positive integer greater than 0.

[0029] The first recognition unit is used to input the N spectrogram image segments into a pre-constructed speaker recognition model to identify the target speaker's target voiceprint feature vector; the speaker recognition model is trained by simultaneously using acoustic features under the original feature domain and the spectral augmentation feature domain, combined with cross-domain feature information.

[0030] The second recognition unit is used to recognize the target speaker based on the target voiceprint feature vector and obtain the recognition result of the target speaker.

[0031] In one possible implementation, the first identification unit includes:

[0032] The first input subunit is used to input the N spectrogram image segments sequentially into the pre-constructed speaker recognition model to obtain N voiceprint feature vectors.

[0033] The calculation subunit is used to perform a weighted average calculation on the N voiceprint feature vectors and use the calculation result as the target voiceprint feature vector of the target speaker.

[0034] In one possible implementation, the device further includes:

[0035] The training unit is used to perform cross-domain training on the initial speaker recognition model using sample speech and a preset target loss function to obtain the trained speaker recognition model.

[0036] The update unit is used to perform verification update training on the trained speaker recognition model using verification speech to obtain the updated speaker recognition model.

[0037] In one possible implementation, the initial speaker recognition model is the mainstream voiceprint recognition model, ECAPA neural network.

[0038] In one possible implementation, the training unit includes:

[0039] The first construction subunit is used to extract acoustic features of sample speech, and use the acoustic features of the sample speech to construct a sample acoustic feature spectrogram corresponding to the sample speech, and divide the sample acoustic feature spectrogram into M sample spectrogram image segments according to a preset window length; where M is a positive integer greater than 0.

[0040] The spectral augmentation subunit is used to perform online spectral augmentation processing on the M sample spectrogram image segments to obtain M sample spectrogram image segments after spectral augmentation.

[0041] The second input subunit is used to input the M sample spectrogram image segments and the M sample spectrogram image segments after spectrum augmentation into the initial speaker recognition model in pairs to obtain M sample voiceprint feature vectors and M spectrum augmented sample voiceprint feature vectors respectively.

[0042] The second construction subunit is used to construct triplet voiceprint feature vectors in the original acoustic feature domain and triplet voiceprint feature vectors in the spectral augmentation feature domain using the M sample voiceprint feature vectors and the M spectral augmentation sample voiceprint feature vectors; and to construct triplet voiceprint feature vectors in different acoustic feature domains.

[0043] The training subunit is used to train the initial speaker recognition model based on the triplet voiceprint feature vector under the original acoustic feature domain, the triplet voiceprint feature vector under the augmented feature domain, the triplet voiceprint feature vector under different acoustic feature domains, and a preset target loss function, so as to obtain the trained speaker recognition model.

[0044] In one possible implementation, the preset target loss function includes a first target loss function, a second target loss function, and a third target loss function; the first target loss function is used to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vector under the same acoustic feature domain, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector; the second target loss function is used to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vector under different acoustic feature domains, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector; the third target loss function is used to improve the accuracy of the speaker recognition model's recognition result.

[0045] In one possible implementation, the acoustic features of the target speech are Mel-frequency cepstral coefficients (MFCCs) or filter bank features.

[0046] This application also provides a speaker recognition device, including: a processor, a memory, and a system bus;

[0047] The processor and the memory are connected via the system bus;

[0048] The memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform any of the above-described implementations of the speaker recognition method.

[0049] This application also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any of the above-described implementations of the speaker recognition method.

[0050] This application also provides a computer program product, which, when run on a terminal device, causes the terminal device to execute any of the above-described speaker recognition methods.

[0051] This application provides a speaker recognition method, apparatus, storage medium, and device. First, it acquires the target speech to be recognized and extracts its acoustic features. Then, using the acoustic features, it constructs an acoustic feature spectrogram corresponding to the target speech and divides the spectrogram into N spectrogram segments according to a preset window length; where N is a positive integer greater than 0. Next, it inputs the N spectrogram segments into a pre-constructed speaker recognition model to identify the target speaker's voiceprint feature vector. The speaker recognition model is trained using acoustic features from both the original feature domain and the spectral augmentation feature domain, combined with cross-domain feature information. Finally, it identifies the target speaker based on the target voiceprint feature vector, obtaining the speaker recognition result.

[0052] As can be seen, since this application constructs a speaker recognition model by first using acoustic features under the original feature domain and the spectral augmented feature domain, and then combining cross-domain feature information to train the model, it effectively improves the speaker feature extraction effect of the model. Therefore, when using this speaker recognition model to perform speaker recognition on the acoustic feature spectrogram of the target speech, it can effectively improve the accuracy of the extracted speaker feature results, thereby improving the accuracy of the final speaker recognition result. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart illustrating a speaker recognition method provided in an embodiment of this application;

[0055] Figure 2 A schematic diagram illustrating the process of constructing a speaker recognition model provided in this application embodiment;

[0056] Figure 3 Example diagrams of sample spectrogram image segments and sample spectrogram image segments after spectroscopic augmentation provided for embodiments of this application;

[0057] Figure 4 This is a schematic diagram illustrating the composition of a speaker recognition device provided in an embodiment of this application. Detailed Implementation

[0058] With the rapid development of intelligent recognition technology, more and more scenarios require the application of voiceprint recognition technology to identify speakers, such as in the financial sector and smart homes. Traditional general-purpose voiceprint recognition technology is the Total Variability system. This system utilizes a large corpus to train a total variable space covering various environments and channels. Using this space, a segment of speech is mapped into a voiceprint model vector (i-vector) of fixed dimension. The similarity of these voiceprint model vectors is then used to describe whether two segments of speech contain the same voiceprint information. However, when the speech duration is short, this method suffers from insufficient statistical calculations, leading to an unstable voiceprint recognition model and a significant drop in authentication accuracy.

[0059] In recent years, deep learning-based recognition methods have seen some application in response to this issue. These methods combine and analyze low-level features to form abstract high-level attribute descriptions, thereby discovering distributed feature representations of the data. Some researchers use deep neural networks (DNNs) to expand the frames of features and then extract voiceprint features by predicting voiceprint identity labels. While this method fully considers the coherence between features and uses features from multiple frames to jointly predict the speaker's identity label, it is still limited to temporal characteristics and cannot establish a correlation between the time and frequency domains, resulting in less than ideal performance in voiceprint recognition.

[0060] To establish the correlation between the time and frequency domains and improve the performance of voiceprint recognition, a common approach is to use an efficient learning method based on deep convolutional neural networks (CNNs). This method first perceives local features, then integrates the local information at higher levels, and utilizes stacked convolutional networks to extract the interwoven voiceprint information between the time and frequency domains. Compared to the simple total variable factor analysis method mentioned above, CNNs can perform joint analysis of the time and frequency domains, deeply mining voiceprint information in the speech spectrum to obtain more detailed voiceprint feature representations, thus enabling the establishment of a more accurate voiceprint recognition model. However, this CNN-based voiceprint modeling method suffers from acoustic feature domain mismatch during training and use, leading to lower accuracy in the model's recognition results.

[0061] Therefore, how to improve the accuracy of voiceprint recognition results, and thus improve the accuracy of speaker recognition results, is a technical problem that urgently needs to be solved.

[0062] To address the aforementioned shortcomings, this application provides a speaker recognition method. First, the target speech to be recognized is acquired, and its acoustic features are extracted. Then, using these acoustic features, an acoustic feature spectrogram corresponding to the target speech is constructed. The spectrogram is then segmented into N spectrogram image segments according to a preset window length, where N is a positive integer greater than 0. Next, the N spectrogram image segments are input into a pre-constructed speaker recognition model to identify the target speaker's voiceprint feature vector. The speaker recognition model is trained using acoustic features from both the original feature domain and the spectral augmentation feature domain, combined with cross-domain feature information. Finally, the target speaker can be identified based on the target voiceprint feature vector, yielding the speaker recognition result.

[0063] As can be seen, since this application constructs a speaker recognition model by first using acoustic features under the original feature domain and the spectral augmented feature domain, and then combining cross-domain feature information to train the model, it effectively improves the speaker feature extraction effect of the model. Therefore, when using this speaker recognition model to perform speaker recognition on the acoustic feature spectrogram of the target speech, it can effectively improve the accuracy of the extracted speaker feature results, thereby improving the accuracy of the final speaker recognition result.

[0064] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] First embodiment

[0066] See Figure 1 This is a flowchart illustrating a speaker recognition method provided in this embodiment. The method includes the following steps:

[0067] S101: Acquire the target speech to be recognized and extract the acoustic features of the target speech.

[0068] In this embodiment, any speaker requiring identification is defined as the target speaker, and the speech of the target speaker that needs to be identified is defined as the target speech. It should be noted that this embodiment does not limit the language type of the target speech; for example, the target speech can be composed of Chinese or English. Furthermore, this embodiment does not limit the length of the target speech; for example, the target speech can be one sentence or multiple sentences.

[0069] Understandably, the target speech can be obtained through recording or other means as needed. For example, telephone conversations or meeting recordings in people's daily lives can be used as target speech. In order to improve the identification of the target speaker, it is necessary to use existing or future feature extraction methods to extract the acoustic features of the target speech for subsequent steps S102.

[0070] Specifically, when extracting acoustic features from target speech, the target speech is first segmented into frames to obtain a corresponding speech frame sequence. Then, the segmented speech frame sequence is pre-emphasized. Next, the acoustic features of each speech frame are extracted sequentially. These acoustic features refer to feature data used to characterize the acoustic information of the corresponding speech frame, and their dimension is defined as 'd'. The specific content format and dimension value are not limited. A preferred implementation is that the acoustic features of the target speech can be, but are not limited to, Mel-scale Frequency Cepstral Coefficients (MFCC) features or filter bank features; the dimension 'd' can be 64 dimensions, etc.

[0071] It should be noted that the embodiments of this application do not limit the method for extracting acoustic features of the target speech, nor do they limit the specific extraction process. Appropriate extraction methods and corresponding feature extraction operations can be selected according to the actual situation.

[0072] S102: Utilize the acoustic features of the target speech to construct the acoustic feature spectrogram corresponding to the target speech, and divide the acoustic feature spectrogram into N spectrogram image segments according to the preset window length; where N is a positive integer greater than 0.

[0073] In this embodiment, after obtaining the target speech to be identified in step S101 and determining its acoustic features (such as 64-dimensional filterbank features), in order to accurately identify the identity information of the target speaker, it is first necessary to process the acoustic features of the target speech using existing or future spectrogram construction methods to construct an acoustic feature spectrogram corresponding to the target speech. The obtained acoustic feature spectrogram is then divided into N spectrogram segments according to a preset window length (defined as l here), with each segment having a size of l×d, for subsequent step S103. Here, N is a positive integer greater than 0.

[0074] Furthermore, it should be noted that this application does not limit the specific value of the preset window length l, which can be set according to the actual situation and experience, such as setting it to 400 frames.

[0075] S103: Input N spectrogram image segments into a pre-built speaker recognition model to identify the target speaker's target voiceprint feature vector; wherein, the speaker recognition model is trained by simultaneously using acoustic features under the original feature domain and the spectral augmented feature domain, combined with cross-domain feature information.

[0076] In this embodiment, after obtaining the N spectrogram image segments corresponding to the target speech in step S102, in order to effectively improve the accuracy of the recognition results, these N spectrogram image segments can be input one by one into a pre-constructed speaker recognition model to identify the target speaker's voiceprint feature vector when uttering the target speech content, which is then used to execute the subsequent step S104. It should be noted that the specific format of the target voiceprint feature vector can be set according to the actual situation; this embodiment does not limit this. For example, the target voiceprint feature vector can be a 256-dimensional vector, etc.

[0077] Specifically, one possible approach is to divide the acoustic spectrogram of the target speech into N spectrogram segments. These N segments can then be sequentially input into a pre-built speaker recognition model to identify the corresponding N voiceprint feature vectors. A weighted average of these N voiceprint feature vectors is then calculated, and the result is used as the target speaker's target voiceprint feature vector. Alternatively, the average of these N voiceprint feature vectors can be directly calculated as the target speaker's target voiceprint feature vector. The specific calculation formula is as follows:

[0078]

[0079] Where w represents the target speaker's voiceprint feature vector; N represents the total number of spectrogram image segments; and wi represents the voiceprint feature vector corresponding to the i-th spectrogram image segment.

[0080] It should be noted that, compared to frame-level acoustic features (such as filterbank features), the target speaker feature vector represents the sentence-level acoustic information of the target speech. It comprehensively considers the relationship between each speech frame and its context, and can more accurately represent the speech information of the target speech. The speaker recognition model is trained by simultaneously using acoustic features under the original feature domain and the spectral augmented feature domain, combined with cross-domain feature information. Therefore, after inputting N spectrogram segments of the target speech into the speaker recognition model, a more accurate target speaker feature vector representing the individual speech information of the target speech can be extracted. Then, using this target speaker feature vector, the target speaker to which the target speech belongs can be identified in the subsequent step S104 to determine their identity information.

[0081] In one possible implementation of this application, to improve the accuracy of speaker recognition results, a speaker recognition model with better recognition performance is pre-constructed. The specific construction process may include a training process and a verification and update process. The training process refers to using pre-collected sample speech and a preset target loss function to perform cross-domain training on an initial speaker recognition model (which may be, but is not limited to, the mainstream voiceprint recognition model ECAPA (Emphasized Channel Attention, Propagation, and Aggregation) neural network) to obtain a trained speaker recognition model. The specific content of the aforementioned preset target loss function can be limited according to actual circumstances, and this application does not limit it. For example, the preset target loss function may be the first target loss function, the second target loss function, the third target loss function, etc., mentioned later. The verification and update process refers to using pre-collected verification speech to perform verification and update training on the trained speaker recognition model. That is, comparing the recognition results with existing calibration results, and further updating the model parameters based on the differences between the two to obtain an updated speaker recognition model with better recognition performance.

[0082] Next, this embodiment will introduce the construction process of the speaker recognition model, such as... Figure 2 As shown, it illustrates the process of constructing a speaker recognition model provided in this embodiment, which includes the following steps A1-A5:

[0083] Step A1: Extract the acoustic features of the sample speech, and use the acoustic features of the sample speech to construct the sample acoustic feature spectrogram corresponding to the sample speech. Then, according to the preset window length, divide the sample acoustic feature spectrogram into M sample spectrogram image segments; where M is a positive integer greater than 0.

[0084] In this embodiment, a lot of preparatory work is required to build a speaker recognition model. First, a large amount of voice data produced by users when they speak needs to be collected. This can be done by using a microphone array to pick up the voice data. The pickup device can be a tablet computer or a smart hardware device, such as a smart speaker, a TV, or an air conditioner. After collecting a large amount of voice data, it is usually necessary to perform noise reduction processing on it. Then, each piece of noise-reduced voice data collected from each user can be used as a sample voice.

[0085] Furthermore, after obtaining the sample speech, it cannot be directly used to train the speaker recognition model. Instead, a method similar to extracting the acoustic features of the target speech in step S101 is required. For example, a feature extractor can be used to replace the target speech with the sample speech to extract the acoustic features of each sample speech. The dimension of each sample speech is also defined as d. For relevant details, please refer to the introduction of step S101, which will not be repeated here.

[0086] Furthermore, a method similar to that used in step S102 to construct the acoustic feature spectrogram corresponding to the target speech can be adopted. The target speech can be replaced with sample speech to construct the sample acoustic feature spectrogram corresponding to each sample speech. For details, please refer to the introduction of step S102, which will not be repeated here.

[0087] Similarly, according to a preset window length (still denoted by l), the obtained sample speech corresponding to the sample acoustic feature spectrogram can be divided into M sample spectrogram image segments, and these segments can be defined as... Furthermore, the size of each spectrogram segment can still be l×d, used to execute subsequent step A2. Here, M is a positive integer greater than 0, and the value of M can be the same as or different from the value of N. It should be noted that if the window length of the sample speech data is less than l, the original sample speech can be copied several times, then segmented, and the excess parts discarded.

[0088] Furthermore, it should be noted that this application does not limit the specific value of the preset window length used during model training; it can be set according to actual conditions and experience, such as 400 frames. Figure 3 As shown, the upper part of the figure (a) can represent the spectrogram image segment of the i-th sample. The horizontal axis represents the time dimension of the spectrogram image segment of the sample, and the vertical axis represents the feature dimension.

[0089] Step A2: Perform online spectrum augmentation on the M sample spectrogram image segments to obtain the M sample spectrogram image segments after spectrum augmentation.

[0090] In this embodiment, step A1 divides the sample acoustic feature spectrogram into M sample spectrogram image segments. To further improve the model's recognition performance, it is necessary to utilize existing or future online spectral augmentation methods to perform online spectral augmentation on the M sample spectrogram image segments, obtaining M sample spectrogram image segments after spectral augmentation. For example, a rectangular region can be randomly defined in both the time dimension and the acoustic feature dimension, and all acoustic feature parameters within these two regions can be set to 0 to obtain the spectral augmented spectrogram segments, which are then defined as... For example, such as Figure 3 As shown, in the lower part of Figure (b), the two dashed boxes represent rectangular regions randomly defined in the time dimension and the acoustic feature dimension, respectively. Figure (b) represents the i-th sample spectrogram segment after spectral augmentation. Its horizontal axis still represents the time dimension of the spectrum image segment of the sample, and its vertical axis still represents the feature dimension.

[0091] Step A3: Input the M sample spectrogram image segments and the M sample spectrogram image segments after spectroscopic augmentation into the initial speaker recognition model in pairs to obtain the M sample voiceprint feature vectors and the M spectroscopic augmented sample voiceprint feature vectors, respectively.

[0092] In this embodiment, M spectrogram image segments of each sample speech are obtained through steps A1 and A2 respectively. and M spectrogram fragments after spectroscopic augmentation Then, the M spectrogram image segments of each sample speech and the M spectrogram image segments after spectral augmentation can be input in pairs into the initial speaker recognition model to obtain the M sample voiceprint feature vectors and the M spectrogram augmented sample voiceprint feature vectors of each sample speech, such as... Figure 2 As shown, each sample voiceprint feature vector and its corresponding spectral augmented sample voiceprint feature vector are defined as w. original and w specaug For example, the voiceprint feature vector of the i-th sample and its corresponding spectral augmented sample voiceprint feature vector are defined as w i original and w i specaug .

[0093] It should be noted that this application does not limit the specific composition structure of the initial speaker recognition model, which can be set according to the actual situation and empirical values. A preferred implementation is that the initial speaker recognition model can be set as the mainstream voiceprint recognition model ECAPA neural network.

[0094] Step A4: Using M sample voiceprint feature vectors and M spectral augmented sample voiceprint feature vectors, construct triplet voiceprint feature vectors in the original acoustic feature domain and triplet voiceprint feature vectors in the spectral augmented feature domain; and construct triplet voiceprint feature vectors in different acoustic feature domains.

[0095] In this embodiment, after obtaining the M sample voiceprint feature vectors and M spectral augmented sample voiceprint feature vectors corresponding to each sample speech through step A3, in order to construct a speaker recognition model with better recognition performance, it is necessary to select sample voiceprint feature vectors that can be used as anchor voiceprint feature vectors under the original acoustic feature domain from the M sample voiceprint feature vectors and M spectral augmented sample voiceprint feature vectors corresponding to each sample speech, and define them as w. a original Then select the one that matches w a original The voiceprint feature vector of a distant sample within the original acoustic feature domain of the same speaker is taken as its corresponding positive example voiceprint feature vector within that feature domain, and is defined as w. p originalThen select the one that matches w a original The closest sample voiceprint feature vector belonging to the original acoustic feature domain of different sample speakers is used as its corresponding counterexample voiceprint feature vector within that feature domain, and is defined as w. n original To construct the triplet voiceprint feature vector (w) under the original acoustic feature domain. a original ,w p original ,w n original ).

[0096] Similarly, we can select the spectral augmented sample voiceprint feature vectors that can serve as anchor voiceprint feature vectors under the spectral augmentation feature domain, and define them as w. a specaug Then select the one that matches w a specaug The speaker feature vector of a more distant spectral augmentation sample within the same spectral augmentation feature domain is taken as its corresponding positive example speaker feature vector within that feature domain, and is defined as w. p specaug Then select the one that matches w a specaug The closest spectral augmented sample voiceprint feature vector belonging to different speakers within the spectral augmented feature domain is taken as its corresponding counterexample voiceprint feature vector within that feature domain, and is defined as w. n specaug To construct a triplet voiceprint feature vector (w) under the spectral augmentation feature domain. a specaug ,w p specaug ,w n specaug ).

[0097] Similarly, anchor point voiceprint feature vectors, positive example voiceprint feature vectors, and negative example voiceprint feature vectors under different feature domains can be used to construct triplet voiceprint feature vectors under different acoustic feature domains, namely: (w a original ,w p specaug ,w n specaug ) and (w a specaug ,w p original ,w n original ).

[0098] Step A5: Train the initial speaker recognition model based on the triplet voiceprint feature vectors in the original acoustic feature domain, the triplet voiceprint feature vectors in the augmented feature domain, the triplet voiceprint feature vectors in different acoustic feature domains, and the preset target loss function to obtain the trained speaker recognition model.

[0099] In this embodiment, the triplet voiceprint feature vector (w) under the original acoustic feature domain is constructed through step A4. a original ,w p original ,w n original ) and triplet voiceprint feature vectors under spectral augmentation feature domain (w a specaug ,w p specaug ,w n specaug ), and triplet voiceprint feature vectors (w) under different acoustic feature domains. a original ,w p specaug ,w n specaug ) and (w a specaug ,w p original ,w n original After that, the initial speaker recognition model can be trained based on the obtained triplet voiceprint feature vectors and the preset target loss function to obtain the trained speaker recognition model.

[0100] Specifically, the softmax function can be used to calculate separately. Figure 2 The normalized posterior probabilities of the speaker's identity label corresponding to the mid-sample acoustic feature spectrogram image segment and the spectrogram image segment after spectral augmentation are defined as p, respectively. original and p specaug Assuming there are 1000 speakers, then the number is p. original and p specaug Each vector represents a 1000-dimensional vector, where each dimension corresponds to the probability that the sample speaker belongs to a predefined speaker. This prediction result is then compared with the actual predefined speaker identity label to which the sample speaker belongs. Based on the comparison result and the value of the predefined target loss function, the model parameters are trained to obtain the trained speaker recognition model.

[0101] One preferred implementation is that the preset target loss function may include a first target loss function, a second target loss function, and a third target loss function. Thus, the initial speaker recognition model is trained using the obtained triplet voiceprint feature vectors and the first, second, and third target loss functions. During training, the model parameters of the initial speaker recognition model are continuously updated based on the changes in the function values ​​of the first, second, and third target loss functions until the function values ​​of the first, second, and third target loss functions meet the requirements, such as the weighted sum of the three reaching a minimum value with minimal change (basically unchanged), or reaching a preset maximum number of iterations (e.g., 10,000). At this point, the update of the model parameters stops, the training of the speaker recognition model is complete, and the trained speaker recognition model is generated.

[0102] The first objective loss function is used to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vector under the same acoustic feature domain, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector. The specific calculation formula is as follows:

[0103] L s =max(0,Dist(w a original ,w n original )+δ-Dist(w a original ,w p original ))+

[0104] max(0,Dist(w a specaug ,w n specaug )+δ-Dist(w a specaug ,w p specaug ))

[0105] Where max(0,Dist(w) a original ,w n original )+δ-Dist(w a original ,w p original )) represents the first part of the first objective loss function; max(0,Dist(w a specaug ,w n specaug)+δ-Dist(w a specaug ,w p specaug )) represents the second part of the first objective loss function; L s The first objective loss function is represented; Dist(*) represents the cosine distance between two voiceprint feature vectors; δ represents the cosine distance boundary, and its specific value is not limited. It can be determined according to the actual situation and empirical values, such as 0.2.

[0106] The second objective function aims to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vectors under different acoustic feature domains, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector. The specific calculation formula is as follows:

[0107] L d =max(0,Dist(w a original ,w n specaug )+δ-Dist(w a original ,w p specaug ))+

[0108] max(0,Dist(w a specaug ,w n original )+δ-Dist(w a specaug ,w p original ))

[0109] Among them, L d The second objective loss function is represented; Dist(*) represents the cosine distance between the two voiceprint feature vectors; δ represents the cosine distance boundary, and its specific value is not limited. It can be determined according to the actual situation and empirical values, such as 0.2.

[0110] In addition, a third objective function is used to improve the accuracy of the speaker recognition model's recognition results. The specific calculation formula is as follows:

[0111]

[0112] in, This represents the first part of the third objective loss function; This represents the second part of the third objective loss function, such as Figure 2 As shown; L w Represents the third objective loss function; and This represents the predicted posterior probability of the real identity label j (which can be manually labeled) in the original acoustic feature domain and the spectral augmented feature domain.

[0113] An alternative implementation is to also use the first objective function L s Second objective function L d The third objective function L w Weighted summations are performed to construct the total loss function, defined as Loss. This Loss function is used to train the initial speaker recognition model. During training, the model parameters are continuously updated based on changes in the Loss function until the Loss function meets certain requirements, such as reaching a minimum value with minimal variation (remaining essentially unchanged), indicating model convergence. At this point, updating the model parameters stops, completing the speaker recognition model training and generating the pre-trained speaker recognition model. The specific formula for calculating Loss is as follows:

[0114] Loss=α·L s +(1-α)·L d +L w

[0115] Wherein, α represents the preset weighting coefficient, which can be set according to the actual situation and experience. This application does not limit its specific value. For example, α can be set to 0.5, etc.

[0116] After the speaker recognition model is trained through the above steps A1-A5, the trained speaker recognition model can be further validated and updated using verification speech to form a speaker recognition model with better recognition performance. The specific verification process will not be described in detail here.

[0117] Thus, this embodiment avoids the problem of input feature mismatch between the model training and inference stages by simultaneously using acoustic features from both the original and augmented feature domains for model training. Furthermore, it proposes a method for model training based on cross-domain triples, incorporating cross-domain voiceprint feature information during training, thereby improving the model's recognition accuracy.

[0118] S104: Based on the target voiceprint feature vector, identify the target speaker and obtain the identification result of the target speaker.

[0119] In this embodiment, after obtaining the target voiceprint feature vector when the target speaker speaks the target speech in step S103, the target speaker can be further identified by processing the target voiceprint feature vector and based on the processing result.

[0120] Specifically, one possible implementation is that the specific implementation process of this step S104 may include the following steps B1-B2:

[0121] Step B1: Calculate the similarity between the target speaker's target voiceprint feature vector and the preset speaker's preset voiceprint feature vector.

[0122] In this implementation, when it is necessary to confirm the identity of the target speaker in order to determine whether the target speaker is a certain preset speaker, after obtaining the target voiceprint feature vector when the target speaker speaks the target speech through step S103, the similarity between the target speaker's target voiceprint feature vector and the preset voiceprint feature vector of the preset speaker can be further calculated. The specific calculation formula is not limited. The higher the value of the similarity, the more similar the target speaker and the preset speaker are, that is, the greater the possibility that the target speaker and the preset speaker are the same person. Conversely, the lower the value of the similarity, the less similar the target speaker and the preset speaker are, that is, the less likely the target speaker and the preset speaker are the same person.

[0123] Step B2: Determine whether the similarity is higher than the preset threshold. If yes, determine that the target speaker is the preset speaker; otherwise, determine that the target speaker is not the preset speaker.

[0124] In this implementation, after calculating the similarity between the target speaker's voiceprint feature vector and the preset speaker's voiceprint feature vector in step B1, it is further necessary to determine whether the similarity is higher than a preset threshold. If so, the target speaker is determined to be the preset speaker; otherwise, the target speaker is determined not to be the preset speaker.

[0125] The preset threshold refers to a critical value used to determine whether the target speaker and the preset speaker are the same person. The specific value can be set according to actual circumstances, and this application embodiment does not limit this; for example, the preset threshold can be set to 0.8. When the similarity between the target speaker's target representation vector and the preset speaker's preset representation vector exceeds this threshold, it indicates that they are the same person. Conversely, when the similarity between the target speaker's target representation vector and the preset speaker's preset representation vector does not exceed this threshold, it indicates that they are not the same person.

[0126] Alternatively, another optional implementation is that the specific implementation process of this step S104 may also include the following steps C1-C2:

[0127] Step C1: Calculate the K similarities between the target speaker's target voiceprint feature vector and the K preset voiceprint feature vectors of K preset speakers; where K is a positive integer greater than 1.

[0128] In this implementation, when it is necessary to identify the target speaker in order to determine which of the K preset speakers (where K is a positive integer greater than 1) is the target speaker, after obtaining the target voiceprint feature vector when the target speaker speaks the target speech through step S103, the K similarities between the target speaker's target voiceprint feature vector and the K preset voiceprint feature vectors of the K preset speakers can be calculated respectively, so as to execute the subsequent step C2.

[0129] Step C2: Select the maximum similarity from the K similarities and determine the target speaker as the preset speaker corresponding to the maximum similarity.

[0130] In this implementation, after calculating the K similarities between the target speaker's target voiceprint feature vector and the K preset representation vectors of K preset speakers in step C1, the maximum similarity can be selected from these K similarities, and the target speaker is determined to be the preset speaker corresponding to the maximum similarity.

[0131] For example: Suppose there are three preset speakers, A, B, and C, and the similarity between the target speaker's voiceprint feature vector and the preset speaker A's voiceprint feature vector is calculated to be 0.1, the similarity between the target speaker's voiceprint feature vector and the preset speaker B's voiceprint feature vector is 0.89, and the similarity between the target speaker's voiceprint feature vector and the preset speaker C's voiceprint feature vector is 0.21. Then, the highest similarity is determined to be 0.89, and based on this highest similarity of 0.89, the identification result is that the target speaker's identity is the preset speaker B.

[0132] In summary, the speaker recognition method provided in this embodiment first acquires the target speech to be recognized and extracts its acoustic features. Then, using the acoustic features of the target speech, an acoustic feature spectrogram corresponding to the target speech is constructed, and the acoustic feature spectrogram is divided into N spectrogram image segments according to a preset window length; where N is a positive integer greater than 0. Next, the N spectrogram image segments are input into a pre-constructed speaker recognition model to identify the target speaker's target voiceprint feature vector. The speaker recognition model is trained by simultaneously using acoustic features under the original feature domain and the spectral augmentation feature domain, combined with cross-domain feature information. Finally, the target speaker can be identified based on the target voiceprint feature vector to obtain the recognition result of the target speaker.

[0133] As can be seen, since this application constructs a speaker recognition model by first using acoustic features under the original feature domain and the spectral augmented feature domain, and then combining cross-domain feature information to train the model, it effectively improves the speaker feature extraction effect of the model. Therefore, when using this speaker recognition model to perform speaker recognition on the acoustic feature spectrogram of the target speech, it can effectively improve the accuracy of the extracted speaker feature results, thereby improving the accuracy of the final speaker recognition result.

[0134] Second embodiment

[0135] This embodiment will introduce a speaker recognition device; please refer to the above method embodiment for related content.

[0136] See Figure 4 This is a schematic diagram of the composition of a speaker recognition device provided in this embodiment. The device 400 includes:

[0137] The acquisition unit 401 is used to acquire the target speech to be identified and extract the acoustic features of the target speech;

[0138] The segmentation unit 402 is used to construct an acoustic feature spectrogram corresponding to the target speech using the acoustic features of the target speech, and to segment the acoustic feature spectrogram into N spectrogram image segments according to a preset window length; where N is a positive integer greater than 0.

[0139] The first recognition unit 403 is used to input the N spectrogram image segments into a pre-constructed speaker recognition model to identify the target speaker's target voiceprint feature vector; the speaker recognition model is trained by simultaneously using acoustic features under the original feature domain and the spectral augmentation feature domain, combined with cross-domain feature information.

[0140] The second recognition unit 404 is used to recognize the target speaker based on the target voiceprint feature vector and obtain the recognition result of the target speaker.

[0141] In one implementation of this embodiment, the first identification unit 403 includes:

[0142] The first input subunit is used to input the N spectrogram image segments sequentially into the pre-constructed speaker recognition model to obtain N voiceprint feature vectors.

[0143] The calculation subunit is used to perform a weighted average calculation on the N voiceprint feature vectors and use the calculation result as the target voiceprint feature vector of the target speaker.

[0144] In one implementation of this embodiment, the apparatus further includes:

[0145] The training unit is used to perform cross-domain training on the initial speaker recognition model using sample speech and a preset target loss function to obtain the trained speaker recognition model.

[0146] The update unit is used to perform verification update training on the trained speaker recognition model using verification speech to obtain the updated speaker recognition model.

[0147] In one implementation of this embodiment, the initial speaker recognition model is the mainstream voiceprint recognition model ECAPA neural network.

[0148] In one implementation of this embodiment, the training unit includes:

[0149] The first construction subunit is used to extract acoustic features of sample speech, and use the acoustic features of the sample speech to construct a sample acoustic feature spectrogram corresponding to the sample speech, and divide the sample acoustic feature spectrogram into M sample spectrogram image segments according to a preset window length; where M is a positive integer greater than 0.

[0150] The spectral augmentation subunit is used to perform online spectral augmentation processing on the M sample spectrogram image segments to obtain M sample spectrogram image segments after spectral augmentation.

[0151] The second input subunit is used to input the M sample spectrogram image segments and the M sample spectrogram image segments after spectrum augmentation into the initial speaker recognition model in pairs to obtain M sample voiceprint feature vectors and M spectrum augmented sample voiceprint feature vectors respectively.

[0152] The second construction subunit is used to construct triplet voiceprint feature vectors in the original acoustic feature domain and triplet voiceprint feature vectors in the spectral augmentation feature domain using the M sample voiceprint feature vectors and the M spectral augmentation sample voiceprint feature vectors; and to construct triplet voiceprint feature vectors in different acoustic feature domains.

[0153] The training subunit is used to train the initial speaker recognition model based on the triplet voiceprint feature vector under the original acoustic feature domain, the triplet voiceprint feature vector under the augmented feature domain, the triplet voiceprint feature vector under different acoustic feature domains, and a preset target loss function, so as to obtain the trained speaker recognition model.

[0154] In one implementation of this embodiment, the preset target loss function includes a first target loss function, a second target loss function, and a third target loss function. The first target loss function is used to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vector under the same acoustic feature domain, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector. The second target loss function is used to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vector under different acoustic feature domains, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector. The third target loss function is used to improve the accuracy of the speaker recognition model's recognition result.

[0155] In one implementation of this embodiment, the acoustic features of the target speech are Mel-frequency cepstral coefficients (MFCC) or filter bank features.

[0156] Furthermore, embodiments of this application also provide a speaker recognition device, including: a processor, a memory, and a system bus;

[0157] The processor and the memory are connected via the system bus;

[0158] The memory is used to store one or more programs, the one or more programs including instructions that, when executed by the processor, cause the processor to perform any of the above-described implementations of the speaker recognition method.

[0159] Furthermore, embodiments of this application also provide a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any of the above-described implementations of the speaker recognition method.

[0160] Furthermore, this application embodiment also provides a computer program product, which, when run on a terminal device, causes the terminal device to execute any of the above-described implementation methods of the speaker recognition method.

[0161] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0162] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0163] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0164] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A speaker recognition method, characterized in that, include: Acquire the target speech to be identified and extract the acoustic features of the target speech; Using the acoustic features of the target speech, an acoustic feature spectrogram corresponding to the target speech is constructed, and the acoustic feature spectrogram is divided into N spectrogram image segments according to a preset window length; where N is a positive integer greater than 0. The N spectrogram image segments are input into a pre-built speaker recognition model to identify the target speaker's voiceprint feature vector; The speaker recognition model is trained by simultaneously using acoustic features from the original feature domain and the spectral augmented feature domain, combined with cross-domain feature information. Based on the target voiceprint feature vector, the target speaker is identified to obtain the identification result of the target speaker.

2. The method according to claim 1, characterized in that, The step of inputting the N spectrogram image segments into a pre-built speaker recognition model to identify the target speaker's target voiceprint feature vector includes: The N spectrogram image segments are sequentially input into a pre-built speaker recognition model to obtain N voiceprint feature vectors; The N voiceprint feature vectors are weighted and averaged, and the result is used as the target voiceprint feature vector of the target speaker.

3. The method according to claim 1, characterized in that, The method further includes: Using sample speech and a preset target loss function, the initial speaker recognition model is trained across domains to obtain the trained speaker recognition model. The trained speaker recognition model is validated and updated using the validation speech to obtain the updated speaker recognition model.

4. The method according to claim 3, characterized in that, The initial speaker recognition model is the mainstream voiceprint recognition model, ECAPA neural network.

5. The method according to claim 3, characterized in that, The step of using sample speech and a preset target loss function to perform cross-domain training on the initial speaker recognition model to obtain the trained speaker recognition model includes: The acoustic features of the sample speech are extracted, and the acoustic features of the sample speech are used to construct the sample acoustic feature spectrogram corresponding to the sample speech. The sample acoustic feature spectrogram is then divided into M sample spectrogram image segments according to a preset window length; where M is a positive integer greater than 0. The M sample spectrogram image segments are subjected to online spectrogram augmentation to obtain M sample spectrogram image segments after spectrogram augmentation. The M sample spectrogram image segments and the M sample spectrogram image segments after spectral augmentation are input into the initial speaker recognition model in pairs to obtain M sample voiceprint feature vectors and M spectral augmented sample voiceprint feature vectors, respectively. Using the M sample voiceprint feature vectors and the M spectral augmented sample voiceprint feature vectors, construct triplet voiceprint feature vectors in the original acoustic feature domain and triplet voiceprint feature vectors in the spectral augmented feature domain; and construct triplet voiceprint feature vectors in different acoustic feature domains. The initial speaker recognition model is trained based on the triplet voiceprint feature vectors in the original acoustic feature domain, the triplet voiceprint feature vectors in the augmented feature domain, the triplet voiceprint feature vectors in different acoustic feature domains, and the preset target loss function, to obtain the trained speaker recognition model.

6. The method according to claim 5, characterized in that, The preset target loss function includes a first target loss function, a second target loss function, and a third target loss function; the first target loss function is used to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vector under the same acoustic feature domain, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector; the second target loss function is used to narrow the distance between the anchor point voiceprint feature vector and the positive example voiceprint feature vector in the triplet voiceprint feature vector under different acoustic feature domains, and to widen the distance between the anchor point voiceprint feature vector and the negative example voiceprint feature vector; the third target loss function is used to improve the accuracy of the speaker recognition model's recognition results.

7. The method according to any one of claims 1-6, characterized in that, The acoustic features of the target speech are Mel-frequency cepstral coefficients (MFCC) or filter bank features.

8. A speaker recognition device, characterized in that, include: An acquisition unit is used to acquire the target speech to be identified and extract the acoustic features of the target speech; The segmentation unit is used to construct an acoustic feature spectrogram corresponding to the target speech using the acoustic features of the target speech, and to segment the acoustic feature spectrogram into N spectrogram image segments according to a preset window length; where N is a positive integer greater than 0. The first recognition unit is used to input the N spectrogram image segments into a pre-built speaker recognition model to identify the target speaker's target voiceprint feature vector. The speaker recognition model is trained by simultaneously using acoustic features from the original feature domain and the spectral augmented feature domain, combined with cross-domain feature information. The second recognition unit is used to recognize the target speaker based on the target voiceprint feature vector and obtain the recognition result of the target speaker.

9. A speaker recognition device, characterized in that, include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, the one or more programs including instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the method described in any one of claims 1-7.

Images