Model training method, voiceprint feature extraction method and its equipment, and program product

Abstract

The model training method, voiceprint feature extraction method and device thereof, and program product provided by the present disclosure relate to a model training method, voiceprint feature extraction method and device thereof, and program product. The technical solution includes: obtaining a first model, a first subframe, and a target subframe applied in a first scene, wherein the first audio has annotation information; extracting the first spectrum feature of the first subframe, and extracting the target spectrum feature of the target subframe; training the first model according to the first spectrum feature of the first subframe, the annotation information of the first audio, the target spectrum feature of the target subframe, and the target audio to which the target subframe belongs, to obtain a target model. In this embodiment, the information of the target audio to which the subframe belongs is used as the annotation information of the subframe, so that the first model can be trained using the first audio and target audio with annotation information to obtain a target model that can recognize the voiceprint features of the audio in the target scene.

Description

Technical Field

[0001] The present disclosure relates to speech technology and deep learning technology in artificial intelligence technology, and in particular to a model training method, a voiceprint feature extraction method, and its equipment and program product. Background Art

[0002] Currently, voiceprint recognition technology has been applied in many scenarios. By performing voiceprint recognition on audio, the identity of the speaker in the audio can be determined. Generally, the model used for voiceprint recognition can be obtained through model training.

[0003] In related technologies, voiceprint recognition models can be applied in a variety of scenarios. To reduce model training costs, existing voiceprint recognition models can be adjusted so that they can be applied in the target scenario. For example, a voiceprint recognition model that can be applied in the insurance field can be optimized and trained using banking-related data to obtain a voiceprint recognition model that can be applied in banking services.

[0004] However, before optimizing and training the model, it is necessary to collect business data for the target scenario and use this business data to train the existing model. However, this approach requires collecting a large amount of business data for the target scenario and labeling this business data before training the existing model. Therefore, this approach is time-consuming and costly. Summary of the Invention

[0005] The present disclosure provides a model training method, a voiceprint feature extraction method, and its equipment and program product to solve the problems of long cycle and high cost in cross-scenario training of existing models in related technologies.

[0006] According to a first aspect of the present disclosure, a model training method is provided, comprising:

[0007] Obtaining a first model, a first subframe, and a target subframe applied to a first scene, wherein the first subframe is obtained by framing first audio applied to the first scene, and the target subframe is obtained by framing target audio applied to a target scene; wherein the first audio has annotation information;

[0008] extracting a first spectrum feature of the first subframe and extracting a target spectrum feature of the target subframe;

[0009] The first model is trained according to the first spectrum feature of the first subframe, the labeling information of the first audio, the target spectrum feature of the target subframe, and the target audio to which the target subframe belongs to obtain a target model.

[0010] According to a second aspect of the present disclosure, a method for extracting voiceprint features is provided, comprising:

[0011] Acquire audio data to be recognized, and extract frequency spectrum features of the audio data;

[0012] The spectral features are input into a target model to obtain the voiceprint features of the audio data; the target model is trained by the method described in the first aspect.

[0013] According to a third aspect of the present disclosure, a model training device is provided, comprising:

[0014] An acquisition unit, configured to acquire a first model, a first subframe, and a target subframe applied to a first scene, wherein the first subframe is obtained by performing frame processing on a first audio applied to the first scene, and the target subframe is obtained by performing frame processing on a target audio applied to a target scene; wherein the first audio has annotation information;

[0015] an extraction unit, configured to extract a first spectrum feature of the first subframe and a target spectrum feature of the target subframe;

[0016] A training unit is configured to train the first model according to the first spectral feature of the first subframe, the annotation information of the first audio, the target spectral feature of the target subframe, and the target audio to which the target subframe belongs, to obtain a target model.

[0017] According to a fourth aspect of the present disclosure, a device for extracting voiceprint features is provided, comprising:

[0018] An audio data acquisition unit, configured to acquire audio data to be recognized and extract frequency spectrum features of the audio data;

[0019] A voiceprint feature extraction unit is used to input the spectral features into a preset target model to obtain the voiceprint features of the audio data; the target model is trained by the device described in the third aspect.

[0020] According to a fifth aspect of the present disclosure, there is provided an electronic device, including:

[0021] at least one processor; and

[0022] a memory communicatively connected to the at least one processor; wherein,

[0023] The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the method according to the first aspect or the second aspect.

[0024] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method as described in the first aspect or the second aspect.

[0025] According to the seventh aspect of the present disclosure, a computer program product is provided, comprising: a computer program, wherein the computer program is stored in a readable storage medium, at least one processor of an electronic device can read the computer program from the readable storage medium, and the at least one processor executes the computer program so that the electronic device executes the method described in the first aspect or the second aspect.

[0026] In the model training method, voiceprint feature extraction method, and device and program product provided by the present disclosure, since the target audio does not have labeling information, the target audio can be split into multiple subframes, and the information of the target audio to which the subframe belongs can be used as the labeling information of the subframe, so that the first model can be trained using the first audio with labeling information and the target audio to obtain a target model that can recognize the voiceprint features of the audio in the target scene.

[0027] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings are provided to facilitate a better understanding of the present invention and do not constitute a limitation of the present disclosure.

[0029] Figure 1 This is a schematic diagram showing an exemplary embodiment of performing voiceprint recognition on audio data;

[0030] Figure 2 A flowchart of a model training method according to an exemplary embodiment of the present disclosure is shown;

[0031] Figure 3 This is a schematic diagram illustrating frame processing of audio according to an exemplary embodiment of the present disclosure;

[0032] Figure 4 A flowchart of a model training method according to another exemplary embodiment of the present disclosure is shown;

[0033] Figure 5 A schematic flow chart of a method for extracting voiceprint features according to an exemplary embodiment of the present disclosure;

[0034] Figure 6 This is a schematic structural diagram of a model training device according to an exemplary embodiment of the present disclosure;

[0035] Figure 7 This is a structural diagram of a model training device according to another exemplary embodiment of the present disclosure;

[0036] Figure 8 This is a schematic structural diagram of a device for extracting voiceprint features according to an exemplary embodiment of the present disclosure;

[0037] Figure 9 is a block diagram of an electronic device for implementing the method according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0038] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0039] Currently, there is a need to identify user identities in many application scenarios. One implementation method is to use voiceprint recognition technology to identify user identities. For example, in a banking business scenario, audio data spoken by the user can be collected, and the voiceprint characteristics of the audio data can be identified using a preset model. The voiceprint characteristics are then compared with the pre-stored user voiceprint characteristics to identify the user's identity.

[0040] Figure 1 The figure is a schematic diagram showing an exemplary embodiment of performing voiceprint recognition on audio data.

[0041] like Figure 1 As shown, a piece of audio data is obtained, and the front-end features of the audio data can be extracted, such as MFCC (Mel-Frequency Cepstral Coefficients), PLP (Perceptual Linear Predictive, perceptual linear prediction coefficient), Fbank (Mel-Frequency Analysis, an evaluation index for speech recognition), or audio features extracted based on FFT (Fast Fourier Transform).

[0042] The front-end features are then input into a pre-trained voiceprint recognition model, which processes the front-end features to extract the speaker's voiceprint features. Voiceprint recognition models can be, for example, GMM, DNN, CNN, ResNet, SincNet, and other structures.

[0043] The back-end classification model (such as COS, LDA, PLDA, etc.) is then used to classify the voiceprint features to identify the speaker.

[0044] Generally, you can train a model for a specific scenario to obtain a model suitable for that scenario. However, this approach is costly. Therefore, you can also optimize and train an existing model for the first scenario to obtain a model suitable for the target scenario. For example, you can further train a model used in banking to obtain a model suitable for insurance.

[0045] In one implementation, a voiceprint recognition model can be optimized and trained to extract voiceprint features. However, this approach requires collecting a large amount of data from the target scenario, labeling the data with a certain amount of manpower, and retraining the voiceprint recognition model to obtain a voiceprint recognition model that can be applied to the target scenario. However, this approach requires a large amount of data to be prepared in advance, and requires a lot of manpower to label the data, resulting in a long training cycle and low training efficiency.

[0046] In another implementation, the backend classification model can be optimized and trained. A certain amount of data from the target scenario can be collected, annotated with human effort, and the backend classification model can be retrained. Because the backend classification model has a simpler structure than the voiceprint recognition model, this implementation does not require updating the more complex voiceprint recognition model; only retraining the backend classification model is required. This reduces the complexity of training, but in this solution, the voiceprint recognition model cannot accurately identify the voiceprint characteristics of the target scenario.

[0047] To address the aforementioned technical issues, the solution provided in this disclosure utilizes the first audio and its annotation information used to train the first scenario, along with the target audio used in the target scenario, to jointly train the existing first model to obtain a target model. This approach allows the target model to draw on the voiceprint recognition capabilities of the first model, and by training the model with the target audio, it can be applied to the target scenario.

[0048] Figure 2 A flowchart of a model training method according to an exemplary embodiment of the present disclosure is shown.

[0049] like Figure 2 As shown, the model training method provided by the present disclosure includes:

[0050] Step 201: Obtain a first model, a first subframe, and a target subframe applied in a first scene, wherein the first subframe is obtained by framing the first audio applied in the first scene, and the target subframe is obtained by framing the target audio applied in the target scene; wherein the first audio has annotation information.

[0051] The method provided by the present disclosure is performed by an electronic device with computing capabilities, such as a computer. The computer can be used to optimize and train the existing first model to obtain a target model.

[0052] Specifically, a first model may be pre-set, and the first model is a model used in a first scenario, specifically a model for recognizing voiceprint features. For example, the first model may be a model used in banking services for recognizing voiceprint features.

[0053] In actual application, the frequency spectrum features of the audio data can be input into the first model, so that the voiceprint features of the audio data can be identified by the first model.

[0054] Furthermore, the first model can be optimized and trained to obtain a model for use in a target scenario, specifically a model for identifying voiceprint features. For example, the model can be used in insurance business to identify voiceprint features.

[0055] In actual application, the electronic device can also obtain a first subframe, which is obtained by framing the first audio. The first audio is audio data used in the first scenario. For example, the first model is a model used in banking business to identify voiceprint features, then the obtained first audio is audio data in banking business.

[0056] In an optional implementation, the first audio may be the first audio used to train the first model, and the first audio has annotation information.

[0057] Among them, the electronic device can also obtain a target subframe, which is obtained by framing the target audio. The target audio is audio data used in the target scene. For example, if the target model is a model used in the insurance business to identify voiceprint features, then the obtained target audio is audio data in the insurance business.

[0058] Specifically, the target audio does not have labeling information, that is, in the solution of the present disclosure, the target audio does not need to be labeled before model training.

[0059] Figure 3 The figure is a schematic diagram showing frame processing of audio according to an exemplary embodiment of the present disclosure.

[0060] Furthermore, the first audio may be subjected to frame processing according to this frame processing method, and the target audio may also be subjected to frame processing using this frame processing method.

[0061] like Figure 3 As shown, for one piece of audio data 31 , a plurality of subframes 32 with a length of 25 ms can be determined in the audio data 31 with a step length of 10 ms.

[0062] Step 202: extract the first spectrum feature of the first subframe and extract the target spectrum feature of the target subframe.

[0063] In one implementation, for each first subframe, the electronic device can extract its first spectrum feature, and for each target subframe, the electronic device can extract its target spectrum feature.

[0064] In another embodiment, when training the target model, a batch of first subframes and target subframes may be obtained, and spectral features may be extracted for the obtained first subframes and target subframes. For example, m first subframes and n target subframes may be obtained, and spectral features may be extracted for each of the obtained subframes.

[0065] In another embodiment, when training the target model, a batch of first audio and target audio can be obtained, each of which can be framed and processed, and the spectral features of each subframe can be extracted. For example, m first audio and n target audio can be obtained, and each of these can be framed and processed to obtain multiple first subframes and target subframes. The first spectral features of each first subframe and the target spectral features of each target subframe can then be extracted.

[0066] The first audio feature and the target spectrum feature are features of the same type, such as MFCC, PLP, Fbank, or features extracted based on FFT.

[0067] Among them, a method for extracting spectral features can be pre-set, such as a method for extracting MFCC of audio data, by which MFCC features of each subframe are extracted. A method for extracting PLP of audio data can also be set, by which PLP features of each subframe are extracted.

[0068] Step 203 : Training the first model according to the first spectrum feature of the first subframe, the annotation information of the first audio, the target spectrum feature of the target subframe, and the target audio to which the target subframe belongs, to obtain a target model.

[0069] Specifically, the first model may be jointly trained using the obtained first spectrum features of the first subframe, the annotation information of the first audio, the target spectrum features of the target subframe, and the target audio to which the target subframe belongs, thereby obtaining a target model.

[0070] Furthermore, the first spectrum feature of the first subframe can be input into the first model, and the first model can perform recognition processing on the first spectrum feature to obtain a recognition result of the first subframe. The recognition result can be the identity of the speaker in the first subframe.

[0071] For example, there are 1000 first audios prepared in advance, and the speaker of each audio is the same user. Therefore, audio data of 300 people can be recorded as the first audio.

[0072] Because the first audio has annotated information, 300 recognition results can be set in the fully connected layer, so that the first model can output the probability that the first subframe belongs to each recognition result. The recognition results and annotated information of the first subframe can be used to adjust the parameters of the first model, thereby training the first model.

[0073] Since the target audio does not have annotated information, it is impossible to pre-set classification information through the fully connected layer. Therefore, the first model cannot output the recognition result of the target subframe. The target feature vector of the target subframe can be extracted by the first model and used to train the first model. The target feature vector extracted by the first model is the voiceprint feature of the target subframe.

[0074] Specifically, a target audio track has only one speaker, so this characteristic can be used to measure the first model's recognition effectiveness for the target subframe. If two target subframes belong to the same target audio track, the target feature vectors of the two target subframes should be relatively close. If the two target subframes belong to different target audio tracks, the target feature vectors of the two target subframes should be dissimilar. Based on this, the parameters of the first model can be adjusted based on the target feature vectors of the target subframes and the target audio tracks to which the target subframes belong, thereby training the first model.

[0075] Furthermore, the first model can be gradually optimized through multiple iterations, and when a preset condition is met, the iteration can be stopped to obtain a target model, which can be applied in a target scene to identify the voiceprint features of audio data in the target scene.

[0076] For example, when the number of iterations reaches a preset number, the model training can be stopped. For another example, when the model has good recognition effects on the first subframe and the target subframe, the model training can also be stopped.

[0077] The model training method provided by the present disclosure includes: obtaining a first model, a first subframe, and a target subframe applied in a first scene, wherein the first subframe is obtained by framing the first audio applied in the first scene, and the target subframe is obtained by framing the target audio applied in the target scene; wherein the first audio has annotation information; extracting the first spectrum feature of the first subframe, and extracting the target spectrum feature of the target subframe; training the first model according to the first spectrum feature of the first subframe, the annotation information of the first audio, the target spectrum feature of the target subframe, and the target audio to which the target subframe belongs, to obtain a target model. In this embodiment, since the target audio does not have annotation information, the target audio can be split into multiple subframes, and the information of the target audio to which the subframe belongs can be used as the annotation information of the subframe, so that the first model can be trained using the first audio and target audio with annotation information to obtain a target model that can recognize the voiceprint features of the audio in the target scene.

[0078] Figure 4 A flowchart of a model training method according to another exemplary embodiment of the present disclosure is shown.

[0079] like Figure 4 As shown, the model training method provided by the present disclosure includes:

[0080] Step 401 : Acquire a first subframe, where the first subframe is obtained by performing frame processing on a first audio applied to a first scene; wherein the first audio has annotation information.

[0081] The method of obtaining the first subframe in step 401 is similar to that in step 201 and will not be described in detail.

[0082] Step 402 : extracting a first spectrum feature of the first subframe, and using the first spectrum feature of the first subframe and the annotation information of the first audio to train a preset model to obtain a first model.

[0083] The first spectrum feature can be extracted for any first subframe based on the following method:

[0084] For any first subframe, the initial spectral features of the first subframe can be determined. For example, an algorithm for extracting spectral features can be pre-set to extract the initial spectral features of each first subframe. For example, features such as MFCC, PLP, and Fbank can be used. Specifically, 40-dimensional initial spectral features of the first subframe can be extracted.

[0085] The associated first subframes of the first subframe are acquired, where the associated first subframes of the first subframe include: a preset number of first subframes located before the first subframe, and / or a preset number of first subframes located after the first subframe.

[0086] Specifically, the electronic device may determine the associated first subframe of the first subframe based on the position of the first subframe in the first audio. For example, a first preset number of first subframes before the first subframe and / or a first preset number of first subframes after the first subframe may be used as the associated first subframes of the first subframe.

[0087] For example, when splitting the first audio, 0-20ms can be used as the first first subframe, 10-30ms as the second first subframe, and 20-40ms as the third first subframe. When the preset number is 1, for the second first subframe, the first first subframe and the third first subframe can be used as its associated first subframes. For the first first subframe, the second first subframe can be used as its associated first subframe. For the third first subframe, the second first subframe and the fourth first subframe can be used as its associated first subframes.

[0088] The characteristic mean of the first subframe may also be determined according to the initial spectrum characteristics of the first subframe associated with the first subframe.

[0089] Furthermore, for one of the first subframes, a characteristic mean of the first subframes may be determined based on the initial spectrum characteristics of the associated first subframes. For example, for the first first subframe, a characteristic mean may be determined based on the initial spectrum characteristics of the associated first subframes.

[0090] In practical applications, the feature mean can be calculated using the initial spectrum features of each associated first subframe. For example, if the initial spectrum feature is a 40-dimensional feature, a mean can be calculated for each dimension to obtain the 40-dimensional feature mean.

[0091] A difference between the initial spectrum feature of the first subframe and the feature mean of the first subframe is determined as the first spectrum feature of the first subframe.

[0092] The electronic device may calculate the difference between the initial spectrum feature of the first subframe and the feature mean to obtain the first spectrum feature of the first subframe. For example, the electronic device may use the 40-dimensional features in the initial spectrum feature to subtract the corresponding 40-dimensional features in the feature mean to obtain the first spectrum feature of the first subframe.

[0093] In this embodiment, by removing the feature mean associated with the first subframe from the initial spectrum feature of the first subframe, noise data in the first subframe can be removed, thereby obtaining a more accurate first spectrum feature.

[0094] The electronic device may use the acquired first subframe and the annotation information of the first audio to which the first subframe belongs to train a preset model, thereby obtaining a first model applied to the first scenario.

[0095] Specifically, the preset model can be composed of multiple layers of TDNN (time delay neural network) and several layers of ResNet. The input data of this preset model is the spectral characteristics of the audio data, and the output recognition result is the probability that the user in the audio data is the respective speaker. For example, if the first audio is recorded audio data of 300 users, the recognition result is the probability that the user in the input data is each of these 300 people.

[0096] Furthermore, the speaker embedding is extracted from the second-to-last hidden layer of the pre-set model. This embedding is used to characterize the user's voiceprint. The final layer of the pre-set model determines the recognition result based on this embedding. When the trained model is applied, the embedding outputs the audio data's feature vector, which is then used to extract the user's voiceprint.

[0097] In actual application, the electronic device can construct a loss function based on the recognition result of the first subframe output by the preset model and the annotation information of the first audio to which the first subframe belongs, and then train the preset model based on the loss function. Through multiple iterative training, the first model can be obtained.

[0098] The source data is the relevant data in the first scene, so the first model obtained is also applied to the first scene. The solution provided by the present disclosure can further optimize and train the first model to obtain a target model applied to the target scene.

[0099] In this implementation, the source data and its annotated information can be used to train a first model, thereby obtaining a first model with voiceprint recognition capabilities. This first model can then be further optimized, and the optimized model can draw on the voiceprint recognition capabilities of the first model. This method of optimizing the training of the first model can reduce the training cost of the target model.

[0100] Step 403 : Acquire a target subframe, where the target subframe is obtained by performing frame processing on the target audio used in the target scene.

[0101] Step 403 is similar to the method of obtaining the target subframe in step 202 and will not be described in detail.

[0102] For any target subframe, steps 404 to 407 may be performed to extract the target spectrum features of the target subframe.

[0103] Step 404: for any target subframe, determine the initial spectrum characteristics of the target subframe.

[0104] Step 405 : Acquire target subframes associated with the target subframe, where the target subframes associated with the target subframe include: a third preset number of target subframes located before the target subframe, and / or a fourth preset number of target subframes located after the target subframe.

[0105] Step 406: Determine a characteristic mean of the target subframe according to the initial spectrum characteristics of the target subframe associated with the target subframe.

[0106] Step 407 : Determine the difference between the initial spectrum feature of the target subframe and the feature mean of the target subframe as the target spectrum feature of the target subframe.

[0107] The implementation of steps 404 to 407 is similar to the method of extracting the first spectrum feature of the first subframe, and will not be repeated here.

[0108] In this embodiment, by removing the feature mean of the associated target subframe from the initial spectrum feature of the target subframe, noise data in the target subframe can be removed, thereby obtaining a more accurate target spectrum feature.

[0109] In one embodiment, steps 408-409 below may be repeatedly performed until a preset training stop condition is met. In another embodiment, steps 401-409 may be repeatedly performed until a preset training stop condition is met.

[0110] Step 408: Input the first spectrum feature and the target spectrum feature into the first model to obtain a recognition result corresponding to the first spectrum feature and a target feature vector corresponding to the target subframe.

[0111] Among them, the electronic device can input the first spectrum characteristics and the target spectrum characteristics into the first model. For example, the first spectrum characteristics obtained through n first audios and the target spectrum characteristics obtained through n target audios can be input into the first model as a batch of training data.

[0112] Specifically, for the first spectrum feature, the first model may output its recognition result, and the recognition result may include the probability of each user, which is used to characterize the probability that the speaker of the first subframe is each user.

[0113] Furthermore, for the target spectrum feature, the first model can output its target feature vector.

[0114] After the first spectral features of the first subframe are input into the first model, the penultimate layer of the first model can extract the source feature vector of the first spectral features. The final layer of the first model can determine the recognition result of the first subframe based on the source feature vector. For example, the probability of the user in the first subframe being each of the 300 users can be determined, such as a 1% probability of being the first user and a 3% probability of being the second user. The final classification result can be determined based on these probabilities. For example, the user with the highest probability can be identified as the speaker of the first subframe.

[0115] Among them, after the target spectrum features of the target subframe are input into the first model, the penultimate layer in the first model can extract the target feature vector of the target spectrum features. Since the target audio does not have labeling information, the classification information cannot be set in advance through the fully connected layer. Therefore, the first model does not classify the target feature vector.

[0116] Specifically, the identifiers of the first spectrum feature and the target spectrum feature can be set in advance, so that the electronic device can distinguish between the first spectrum feature and the target spectrum feature. For the first spectrum feature, the electronic device can use the first model to determine its recognition result, and for the target spectrum feature, the electronic device can use the first model to determine its target feature vector.

[0117] In step 409 , a loss function value is determined using the recognition result of the first spectral feature, the annotation information of the first audio, the target feature vector of the target subframe, and the target audio to which the target subframe belongs, and the loss function value is used to optimize the first model to obtain an optimized first model.

[0118] In actual application, the first audio to which the first subframe belongs has label information, so the label information of the first audio can also be used as the label information of the first subframe. For example, the recognition result of the first subframe and the label information of the first subframe can be used to train the first model.

[0119] Specifically, a target audio track has only one speaker. Therefore, this characteristic can be used to measure the recognition performance of the first model for the target subframe. If two target subframes belong to the same target audio track, the target feature vectors of the two target subframes should be relatively close. If the two target subframes belong to different target audio tracks, the target feature vectors of the two target subframes should be dissimilar. Based on this, the first model can be trained based on the target feature vectors of the target subframes and the target audio tracks to which the target subframes belong.

[0120] Furthermore, the value of the loss function can be determined by combining the recognition result of the first spectral feature, the annotation information of the first audio, the target feature vector of the target subframe, and the target audio to which the target subframe belongs, so as to jointly train the first model using the first audio and the target audio to obtain the target model.

[0121] In practical applications, the first model can be iteratively trained multiple times, and training can be stopped when a preset stopping condition is met to obtain a target model. For example, training can be stopped when the number of training cycles reaches a preset number, or when the value of the loss function is less than a preset value.

[0122] Through this implementation method, the first model can be jointly trained using the first audio and the target audio. Although the target audio has no labeled information, the method provided by the present disclosure can use the characteristic that the speakers of the target subframes belonging to the same target audio are the same user to train the first model using the target audio, thereby obtaining a target model that can be applied to the target scene. At the same time, the target model is trained using the first audio to avoid overfitting of the target model to the data of the target scene.

[0123] The value of the first loss function may be determined according to the recognition result of the first spectrum feature and the labeling information of the first spectrum feature.

[0124] In practical applications, the first audio to which the first subframe belongs has annotated information. Therefore, the annotated information of the first audio can also be used as the annotated information of the first subframe. Therefore, the value of the first loss function can be determined using the recognition result of the first subframe and the annotated information of the first audio to which the first subframe belongs. The value of the first loss function is used to measure the recognition effect of the first model on the first subframe.

[0125] The value of the second loss function can be determined based on the target feature vector of each target subframe and the target audio to which each target subframe belongs; the value of the second loss function is used to characterize the comparison information between the target feature vectors of multiple target subframes.

[0126] Among them, the target audio to which the target subframe belongs does not have labeling information, but if the two target subframes belong to the same target audio, the target feature vectors of the two target subframes should be relatively close, and if the two target subframes belong to different target audios, the target feature vectors of the two target subframes should be dissimilar. Therefore, the target feature vectors of the target subframes belonging to the same target audio and the target feature vectors of the target subframes belonging to different target audios can be used to determine the value of the second loss function. The value of the second loss function can be used to measure the recognition effect of the first model on the target subframe.

[0127] Specifically, the value of the loss function can be determined based on the values ​​of the first loss function and the second loss function. This loss function can measure the overall recognition effect of the model during the training process. The parameters in the model can be adjusted using this loss function to obtain a target model that meets the requirements.

[0128] In an optional embodiment, the difference between the value of the first loss function and the value of the weighted second loss function can be used to determine the value of the loss function of the model. For example, the value of the loss function can be:

[0129] L total =L cla +λL cl

[0130] Among them, L total is the value of the loss function, L cla is the value of the first loss function, L cl is the value of the second loss function, and λ is a preset weighting coefficient. In this embodiment, the user can use λ to adjust the ratio of the value of the first loss function to the value of the second loss function, thereby meeting the needs of multiple cross-scenario model training.

[0131] In this way, the value of the model's loss function can be determined, which is used to measure the model's recognition effect on the first spectral feature and the target spectral feature. This value can then be used to optimize the model. The optimized model can more accurately identify the first spectral feature and the target spectral feature, and through multiple iterations, a target model that can accurately identify the target spectral feature can be obtained.

[0132] Furthermore, when determining the value of the second loss function, the first comparison information can be determined based on the feature vectors of the target subframes belonging to the same target audio. Since the target subframes belong to the same target audio, the target feature vectors of the target subframes should be relatively close. Therefore, if the first comparison information is small, it can be characterized that the model has a good recognition effect on the audio data of the same user. Otherwise, the recognition effect of the model on the audio data of the same user is poor.

[0133] In practical applications, the second comparison information can also be determined based on the feature vectors of target subframes belonging to different target audios. Since the target subframes belong to different target audios, the target feature vectors of these target subframes should differ significantly. Therefore, if the second comparison information is small, it can be indicated that the model has poor recognition performance for audio data from different users. Otherwise, it can be indicated that the model has good recognition performance for audio data from different users.

[0134] Finally, the first and second comparison information can be combined to determine the value of the second loss function. Since the smaller the value of the first comparison information and the larger the value of the second comparison information, the better the recognition effect of the model, the first and second comparison information can be used to determine a value that can measure the recognition effect of the model.

[0135] Through this implementation, even if the target audio does not have labeled information, the value of the second loss function used to measure the model's recognition effect on the target audio can be determined based on the target audio to which the target subframe belongs. Furthermore, without the need to label the target audio, the target audio can be used to optimize and train the existing first model to obtain the target model.

[0136] Figure 5 The figure is a flowchart of a method for extracting voiceprint features according to an exemplary embodiment of the present disclosure.

[0137] like Figure 5 As shown, the voiceprint feature extraction method provided by the present disclosure includes:

[0138] Step 501: Acquire audio data to be recognized and extract frequency spectrum features of the audio data.

[0139] Step 502: Input the spectral features into a preset target model to obtain the voiceprint features of the audio data.

[0140] The method provided in the present disclosure can be applied to electronic devices with computing capabilities, such as computers or mobile phones, and the target model can be set in the electronic device.

[0141] Specifically, a preset spectrum feature extraction algorithm may be used to extract spectrum features from the audio data to be identified, and then the extracted spectrum features may be processed using a target model to obtain voiceprint features of the audio data.

[0142] Furthermore, the target model can be obtained by Figure 2 or Figure 4 The model trained by the method shown can use the voiceprint feature of the audio data output by the penultimate layer of the target model. The voiceprint feature can be, for example, a feature vector.

[0143] In an optional embodiment, if the electronic device is a mobile phone, the electronic device can send the obtained voiceprint feature to a server, and the server compares the voiceprint feature with a preset voiceprint feature to determine the user who matches the voiceprint feature. The preset voiceprint feature is associated with user information.

[0144] In another optional embodiment, if the electronic device is a computer, a voiceprint feature may be pre-set in the computer, and the pre-set voiceprint feature may be associated with user information. The computer may compare the voiceprint feature with the pre-set voiceprint feature to determine a user that matches the voiceprint feature.

[0145] Figure 6 The figure is a schematic structural diagram of a model training device according to an exemplary embodiment of the present disclosure.

[0146] like Figure 6 As shown, the model training device 600 provided by the present disclosure includes:

[0147] A training data acquisition unit 610 is configured to acquire a first model, a first subframe, and a target subframe for use in a first scene, wherein the first subframe is obtained by framing first audio for use in the first scene, and the target subframe is obtained by framing target audio for use in a target scene; wherein the first audio has annotation information;

[0148] a spectrum feature extraction unit 620, configured to extract a first spectrum feature of the first subframe and a target spectrum feature of the target subframe;

[0149] The training unit 630 is configured to train the first model according to the first spectral features of the first subframe, the annotation information of the first audio, the target spectral features of the target subframe, and the target audio to which the target subframe belongs, to obtain a target model.

[0150] In the model training device provided by the present disclosure, since the target audio does not have labeling information, the target audio can be split into multiple subframes, and the information of the target audio to which the subframe belongs can be used as the labeling information of the subframe, so that the first model can be trained using the first audio with labeling information and the target audio to obtain a target model that can recognize the voiceprint features of the audio in the target scene.

[0151] Figure 7 This is a schematic structural diagram of a model training device according to another exemplary embodiment of the present disclosure.

[0152] like Figure 7 As shown, in the model training device 700 provided by the present disclosure, the training data acquisition unit 710 and Figure 6 Similar to the training data acquisition unit 610 in FIG, the spectrum feature extraction unit 720 is similar to Figure 6 Similar to the spectrum feature extraction unit 620 in FIG, the training unit 730 is similar to Figure 6 The training unit 630 in is similar.

[0153] The training unit 730 includes a recognition module 731 and a training module 732:

[0154] The identification module 731 and the training module 732 repeatedly perform the following steps until a preset training stop condition is met:

[0155] The recognition module 731 inputs the first spectrum feature and the target spectrum feature into the first model to obtain a recognition result corresponding to the first spectrum feature and a target feature vector corresponding to the target subframe;

[0156] The training module 732 determines a value of a loss function using the recognition result of the first spectral feature, the labeling information of the first audio, the target feature vector of the target subframe, and the target audio to which the target subframe belongs, and optimizes the first model using the value of the loss function to obtain an optimized first model;

[0157] The optimized first model obtained when the preset training stop condition is met is the target model.

[0158] The training module 732 is further configured to:

[0159] determining a value of a first loss function according to the recognition result of the first spectral feature and the labeling information of the first spectral feature;

[0160] Determining a value of a second loss function according to the target feature vectors of each target subframe and the target audio to which each target subframe belongs; the value of the second loss function is used to represent comparison information between the target feature vectors of the plurality of target subframes;

[0161] The value of the loss function is determined according to the value of the first loss function and the value of the second loss function.

[0162] The training module 732 is further configured to:

[0163] determining first comparison information according to target feature vectors of target subframes belonging to the same target audio;

[0164] determining second comparison information according to target feature vectors of target subframes belonging to different target audios;

[0165] Determine a value of the second loss function based on the first comparison information and the second comparison information.

[0166] The training module 732 is further configured to:

[0167] The difference between the value of the first loss function and the value of the weighted second loss function is determined as the value of the loss function.

[0168] The spectrum feature extraction unit 720 includes a first feature extraction module 721, which is used to:

[0169] For any first subframe, determining an initial spectrum characteristic of the first subframe;

[0170] Acquire associated first subframes of the first subframe, where the associated first subframes include: a first preset number of first subframes located before the first subframe, and / or a second preset number of first subframes located after the first subframe;

[0171] determining a characteristic mean of the first subframe according to an initial spectrum characteristic of the first subframe associated with the first subframe;

[0172] A difference between the initial spectrum feature of the first subframe and the feature mean of the first subframe is determined as the first spectrum feature of the first subframe.

[0173] The spectrum feature extraction unit 720 includes a target feature extraction module 722, which is used to:

[0174] For any target subframe, determining the initial spectrum characteristics of the target subframe;

[0175] Acquire target subframes associated with the target subframe, where the target subframes associated with the target subframe include: a third preset number of target subframes located before the target subframe, and / or a fourth preset number of target subframes located after the target subframe;

[0176] determining a characteristic mean of the target subframe according to an initial spectrum characteristic of a target subframe associated with the target subframe;

[0177] A difference between the initial spectrum feature of the target subframe and the feature mean of the target subframe is determined as the target spectrum feature of the target subframe.

[0178] In the apparatus 700 provided in this embodiment, the training data acquisition unit 710 is further configured to:

[0179] A preset model is trained using the first spectrum feature of the first subframe and the labeled information of the first audio to obtain the first model.

[0180] Figure 8 The figure is a schematic structural diagram of a device for extracting voiceprint features according to an exemplary embodiment of the present disclosure.

[0181] like Figure 8 As shown, the voiceprint feature extraction device 800 provided by the present disclosure includes:

[0182] The audio data acquisition unit 810 is used to acquire the audio data to be recognized and extract the frequency spectrum characteristics of the audio data;

[0183] The voiceprint feature extraction unit 820 is used to input the spectrum feature into a preset target model to obtain the voiceprint feature of the audio data; the target model is Figure 6 or Figure 7 The device is trained by any one of the devices shown.

[0184] The present disclosure provides a model training method, a voiceprint feature extraction method, and its equipment and program products, which are applied to speech technology and deep learning technology in artificial intelligence technology to solve the problems of long cycle and high cost in cross-scenario training of existing models in related technologies.

[0185] It should be noted that the first model and target model in this embodiment are not voiceprint recognition models for a specific user and cannot reflect the personal information of a specific user. It should be noted that the audio data in this embodiment comes from a public dataset.

[0186] In the technical solutions disclosed herein, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0187] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0188] According to an embodiment of the present disclosure, the present disclosure also provides a computer program product, which includes: a computer program, the computer program is stored in a readable storage medium, at least one processor of an electronic device can read the computer program from the readable storage medium, and at least one processor executes the computer program so that the electronic device executes the solution provided by any of the above embodiments.

[0189] Figure 9 A schematic block diagram of an example electronic device 900 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are provided as examples only and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0190] like Figure 9 As shown, the device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. Various programs and data required for the operation of the device 900 can also be stored in the RAM 903. The computing unit 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0191] Various components in the device 900 are connected to the I / O interface 905, including an input unit 906, such as a keyboard, a mouse, etc.; an output unit 907, such as various types of displays, speakers, etc.; a storage unit 908, such as a magnetic disk, an optical disk, etc.; and a communication unit 909, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 909 allows the device 900 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0192] The computing unit 901 can be any general-purpose and / or specialized processing component with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as the model training method or the voiceprint feature extraction method. For example, in some embodiments, the model training method or the voiceprint feature extraction method can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 900 via the ROM 902 and / or the communication unit 909. When the computer program is loaded into the RAM 903 and executed by the computing unit 901, one or more steps of the model training method or the voiceprint feature extraction method described above can be performed. Alternatively, in other embodiments, the computing unit 901 may be configured to execute the model training method or the voiceprint feature extraction method in any other appropriate manner (for example, by means of firmware).

[0193] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0194] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0195] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0196] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0197] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0198] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact via a communication network. This client-server relationship is established by computer programs running on the respective computers, establishing a client-server relationship. The server may be a cloud server, also known as a cloud computing server or cloud host, a host product within the cloud computing service ecosystem that addresses the management difficulties and limited scalability of traditional physical hosts and VPS services ("Virtual Private Servers" or simply "VPS"). The server may also be a server in a distributed system or a server integrated with blockchain.

[0199] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not a limitation herein.

[0200] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A model training method, comprising: Obtaining a first model, a first subframe, and multiple target subframes applied to a first scene, where the first subframe is obtained by framing first audio applied to the first scene, and the multiple target subframes are obtained by framing target audio applied to a target scene; wherein the first audio has annotation information; and the multiple target subframes include subframes belonging to the same target audio and subframes that do not belong to the same target audio; extracting a first spectrum feature of the first subframe and extracting a target spectrum feature of the target subframe; Repeat the following steps until the preset stop training condition is met: Inputting the first spectrum feature and the target spectrum feature into the first model to obtain a recognition result corresponding to the first spectrum feature and a target feature vector corresponding to the target subframe; Determining a value of a loss function using the recognition result of the first spectral feature, the labeling information of the first audio, the target feature vector of the target subframe, and the target audio to which the target subframe belongs, and adjusting parameters in the first model using the value of the loss function to optimize the first model, thereby obtaining an optimized first model; Among them, the optimized first model obtained when the preset stop training condition is met is the target model; the target model is used to obtain the user's voiceprint features.

2. The method according to claim 1, wherein The determining a value of a loss function by using the recognition result of the first spectral feature, the labeling information of the first audio, the target feature vector of the target subframe, and the target audio to which the target subframe belongs includes: Determining a value of a first loss function according to the recognition result of the first spectral feature and the annotation information of the first audio; Determining a value of a second loss function according to the target feature vectors of each target subframe and the target audio to which each target subframe belongs; the value of the second loss function is used to represent comparison information between the target feature vectors of the plurality of target subframes; The value of the loss function is determined according to the value of the first loss function and the value of the second loss function.

3. The method according to claim 2, wherein: The determining a value of the second loss function according to the target feature vector of each target subframe and the target audio to which each target subframe belongs includes: determining first comparison information according to target feature vectors of target subframes belonging to the same target audio; determining second comparison information according to target feature vectors of target subframes belonging to different target audios; Determine a value of the second loss function based on the first comparison information and the second comparison information.

4. The method according to claim 2 or 3, wherein: The determining the value of the loss function according to the value of the first loss function and the value of the second loss function includes: The difference between the value of the first loss function and the value of the weighted second loss function is determined as the value of the loss function.

5. The method according to any one of claims 1 to 3, wherein: The extracting the first spectrum feature of the first subframe includes: For any first subframe, determining an initial spectrum characteristic of the first subframe; Acquire associated first subframes of the first subframe, where the associated first subframes include: a first preset number of first subframes located before the first subframe, and / or a second preset number of first subframes located after the first subframe; determining a characteristic mean of the first subframe according to an initial spectrum characteristic of the first subframe associated with the first subframe; A difference between the initial spectrum feature of the first subframe and the feature mean of the first subframe is determined as the first spectrum feature of the first subframe.

6. The method according to claim 1, wherein The extracting the target spectrum feature of the target subframe includes: For any target subframe, determining an initial spectrum characteristic of the target subframe; Acquire target subframes associated with the target subframe, where the target subframes associated with the target subframe include: a third preset number of target subframes located before the target subframe, and / or a fourth preset number of target subframes located after the target subframe; determining a characteristic mean of the target subframe according to an initial spectrum characteristic of a target subframe associated with the target subframe; A difference between the initial spectrum feature of the target subframe and the feature mean of the target subframe is determined as the target spectrum feature of the target subframe.

7. The method according to any one of claims 1 to 3 and 6, wherein obtaining the first model applied in the first scenario comprises: A preset model is trained using the first spectrum feature of the first subframe and the labeled information of the first audio to obtain the first model.

8. A method for extracting voiceprint features, comprising: Acquire audio data to be recognized, and extract frequency spectrum features of the audio data; Inputting the spectral features into a target model to obtain the voiceprint features of the audio data; The target model is obtained by training using the method described in any one of claims 1 to 7.

9. A model training device comprising: a training data acquisition unit, configured to acquire a first model, a first subframe, and a plurality of target subframes applied to a first scene, wherein the first subframe is obtained by framing a first audio applied to the first scene, and the plurality of target subframes are obtained by framing a target audio applied to a target scene; wherein the first audio has annotation information; and the plurality of target subframes include subframes belonging to the same target audio and subframes that do not belong to the same target audio; a spectrum feature extraction unit, configured to extract a first spectrum feature of the first subframe and a target spectrum feature of the target subframe; Training unit, including recognition module and training module: The recognition module and the training module repeatedly perform the following steps until a preset training stop condition is met: The recognition module inputs the first spectrum feature and the target spectrum feature into the first model to obtain a recognition result corresponding to the first spectrum feature and a target feature vector corresponding to the target subframe; The training module determines a value of a loss function using the recognition result of the first spectral feature, the labeling information of the first audio, the target feature vector of the target subframe, and the target audio to which the target subframe belongs, and adjusts parameters in the first model using the value of the loss function to optimize the first model, thereby obtaining an optimized first model; Among them, the optimized first model obtained when the preset stop training condition is met is the target model; the target model is used to obtain the user's voiceprint features.

10. The device according to claim 9, wherein The training module is also used to: determining a value of a first loss function according to the recognition result of the first spectral feature and the labeling information of the first spectral feature; Determining a value of a second loss function according to the target feature vectors of each target subframe and the target audio to which each target subframe belongs; the value of the second loss function is used to represent comparison information between the target feature vectors of the plurality of target subframes; The value of the loss function is determined according to the value of the first loss function and the value of the second loss function.

11. The device according to claim 10, wherein The training module is also used to: determining first comparison information according to target feature vectors of target subframes belonging to the same target audio; determining second comparison information according to target feature vectors of target subframes belonging to different target audios; Determine a value of the second loss function based on the first comparison information and the second comparison information.

12. The device according to claim 10 or 11, wherein The training module is also used to: The difference between the value of the first loss function and the value of the weighted second loss function is determined as the value of the loss function.

13. The device according to any one of claims 9 to 11, wherein: The spectrum feature extraction unit includes a first feature extraction module, which is used to: For any first subframe, determining an initial spectrum characteristic of the first subframe; Acquire associated first subframes of the first subframe, where the associated first subframes include: a first preset number of first subframes located before the first subframe, and / or a second preset number of first subframes located after the first subframe; determining a characteristic mean of the first subframe according to an initial spectrum characteristic of the first subframe associated with the first subframe; A difference between the initial spectrum feature of the first subframe and the feature mean of the first subframe is determined as the first spectrum feature of the first subframe.

14. The device according to any one of claims 9 to 11, wherein: The spectrum feature extraction unit includes a target feature extraction module, which is used to: For any target subframe, determining the initial spectrum characteristics of the target subframe; Acquire target subframes associated with the target subframe, where the target subframes associated with the target subframe include: a third preset number of target subframes located before the target subframe, and / or a fourth preset number of target subframes located after the target subframe; determining a characteristic mean of the target subframe according to an initial spectrum characteristic of a target subframe associated with the target subframe; A difference between the initial spectrum feature of the target subframe and the feature mean of the target subframe is determined as the target spectrum feature of the target subframe.

15. The apparatus according to any one of claims 9 to 11, wherein the training data acquisition unit is further configured to: A preset model is trained using the first spectrum feature of the first subframe and the labeled information of the first audio to obtain the first model.

16. A device for extracting voiceprint features, comprising: An audio data acquisition unit, configured to acquire audio data to be recognized and extract frequency spectrum features of the audio data; A voiceprint feature extraction unit is used to input the spectral features into a preset target model to obtain the voiceprint features of the audio data; the target model is trained by the device according to any one of claims 9 to 15.

17. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 8.

18. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-8.

19. A computer program product comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 8.

Images