Acoustic biomimetic voiceprint recognition method and device, electronic equipment and storage medium

Abstract

The application provides an auditory bionic voiceprint recognition method and device, electronic equipment and a storage medium. The auditory bionic voiceprint recognition method comprises the following steps: obtaining a to-be-recognized voice signal, and performing cutting processing on the to-be-recognized voice signal to obtain a plurality of to-be-recognized voice segments; inputting each to-be-recognized voice segment into a cochlea bionic filter to obtain a voice frequency feature; inputting the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result; and the voiceprint recognition model is obtained based on a cochlea nucleus bionic network. The cochlea bionic filter and the voiceprint recognition model provided by the application have higher adaptability to noise and complex environments, thereby improving the reliability and efficiency of voiceprint recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biometric recognition, in particular to an auditory biomimetic voiceprint recognition method and device, electronic equipment and storage medium. BACKGROUND

[0002] Biometric recognition technology is a technology that detects and analyzes unique biological characteristics of the human body (such as fingerprints, faces, irises, voiceprints, etc.) to identify identity, with high security and convenience. Voiceprint is a unique biological characteristic that contains characteristic information in human speech signals, and identity authentication is achieved by analyzing the voice characteristics of individuals. With the rapid development of biometric recognition technology, voiceprint recognition has gradually become an important biometric authentication method.

[0003] In recent years, deep learning has made breakthroughs in many fields and has been widely applied. However, existing deep learning-based methods have the following shortcomings in the application of voiceprint recognition: first, poor interpretability, which makes it difficult to trace the decision-making process of the model and reduces the credibility and application security of the system; second, existing deep learning methods usually use filter banks, mel-frequency cepstral coefficients, and spectrograms as acoustic features as model inputs, lacking the ability to directly process raw audio signals. In addition, some end-to-end voiceprint recognition systems fail to apply the characteristics of the speech signal processing pathway in the auditory system of living organisms to the design of neural network model structures, resulting in models that cannot fully utilize these characteristics to enhance the extraction of speech features, affecting their adaptability and performance in practical applications. SUMMARY

[0004] The present application provides an auditory biomimetic voiceprint recognition method, device, electronic equipment and storage medium to solve the defects in the prior art, achieve higher adaptability to noise and complex environments, and improve the reliability and efficiency of voiceprint recognition.

[0005] The present application provides an auditory biomimetic voiceprint recognition method, comprising:

[0006] Obtaining a voice signal to be recognized, and cutting the voice signal to be recognized to obtain a plurality of voice segments to be recognized;

[0007] Inputting each of the voice segments to be recognized into a cochlea biomimetic filter to obtain a voice frequency feature;

[0008] Inputting the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result; wherein the voiceprint recognition model is obtained based on a cochlear nucleus biomimetic network.

[0009] According to the hearing bionic voiceprint recognition method provided by the application, the voiceprint recognition model comprises a first cochlear nucleus bionic network, a full connection network and an output layer connected in sequence.

[0010] The first cochlear nucleus bionic network is configured to extract time sequence features and spatial features based on the voice frequency features, obtain time sequence feature representation based on the time sequence features and the spatial features, and perform weighted fusion processing on the time sequence feature representation and the voice frequency features to obtain voice feature representation.

[0011] The full connection network and the output layer are configured to sequentially process the voice feature representation to obtain the voiceprint recognition result.

[0012] According to the hearing bionic voiceprint recognition method provided by the application, the first cochlear nucleus bionic network comprises a spiral ganglion neuron layer, a star cell layer, an attention mechanism and a cluster cell layer, the attention mechanism is connected with the spiral ganglion neuron layer, the star cell layer and the cluster cell layer respectively, and the spiral ganglion neuron layer is connected with the star cell layer.

[0013] The spiral ganglion neuron layer is configured to perform convolution operation in the time direction based on the voice frequency features to obtain first bionic features.

[0014] The star cell layer is configured to perform convolution operation in the time direction based on the first bionic features to obtain second bionic features.

[0015] The attention mechanism is configured to perform weighted fusion on the voice frequency features and the second bionic features through attention parameters to obtain intermediate features.

[0016] The cluster cell layer is configured to perform convolution operation in the time direction based on the intermediate features to obtain the voice feature representation.

[0017] According to the hearing bionic voiceprint recognition method provided by the application, the output layer comprises a Softmax function or an additive margin Softmax function.

[0018] According to the hearing bionic voiceprint recognition method provided by the application, the voiceprint recognition model further comprises a second cochlear nucleus bionic network, and the second cochlear nucleus bionic network is connected between the first cochlear nucleus bionic network and the full connection network.

[0019] According to the hearing bionic voiceprint recognition method provided by the application, the cochlear bionic filter comprises a filter bank composed of a plurality of filters with different response frequencies; wherein each filter extracts the voice frequency features corresponding to the response frequency from the voice segment to be recognized.

[0020] According to the hearing bionic voiceprint recognition method provided by the application, the training process of the voiceprint recognition model comprises:

[0021] obtaining a training set; wherein the training set comprises a plurality of training samples;

[0022] constructing a loss function based on the training samples;

[0023] inputting the training samples in the training set into a basic voiceprint recognition model to perform iterative training with the training target of minimizing the loss function until the training is completed, so as to obtain the voiceprint recognition model; wherein the basic voiceprint recognition model is obtained based on a cochlear nucleus bionic network.

[0024] The application further provides a hearing bionic voiceprint recognition device, comprising:

[0025] a first obtaining module configured to obtain a to-be-recognized voice signal and perform cutting processing on the to-be-recognized voice signal to obtain a plurality of to-be-recognized voice segments;

[0026] a first input module configured to input each to-be-recognized voice segment into a cochlear bionic filter to obtain a voice frequency feature;

[0027] a second input module configured to input the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result; wherein the voiceprint recognition model is obtained based on a cochlear nucleus bionic network.

[0028] The application further provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the hearing bionic voiceprint recognition method according to any one of the above when executing the program.

[0029] The application further provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the hearing bionic voiceprint recognition method according to any one of the above.

[0030] The application further provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to implement the hearing bionic voiceprint recognition method according to any one of the above.

[0031] The auditory bionic voiceprint recognition method, device, electronic device, and storage medium provided by this invention construct an initial feature set related to the original speech signal by segmenting the speech signal to be recognized into speech segments. These speech segments are then input into a cochlear bionic filter. Based on the characteristics of the cochlear basilar membrane, the cochlear bionic filter simulates the response of hair cells to specific frequencies, directly and efficiently extracting speech frequency features from the original speech waveform. Voiceprint recognition is then performed using the voiceprint recognition model provided by this invention. This achieves greater adaptability to noise and complex environments, thereby improving the reliability and efficiency of voiceprint recognition. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in this invention 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 invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating the auditory bionic voiceprint recognition method provided by the present invention.

[0034] Figure 2 This is one of the structural schematic diagrams of the voiceprint recognition model provided in the embodiments of the present invention.

[0035] Figure 3 This is a schematic diagram of the local regulatory circuit of the ventral cochlear nucleus in an embodiment provided by the present invention.

[0036] Figure 4 This is a schematic diagram of the connection of neurons in the cochlear nucleus biomimetic network provided in the embodiment of the present invention.

[0037] Figure 5 This is the second structural schematic diagram of the voiceprint recognition model provided in the embodiments of the present invention.

[0038] Figure 6 This is the third structural schematic diagram of the voiceprint recognition model provided in the embodiments of the present invention.

[0039] Figure 7 This is a schematic diagram of the structure of the auditory bionic voiceprint recognition device provided by the present invention.

[0040] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0041] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in conjunction with the drawings in the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0042] The technical solutions in the present application will be described below in conjunction with the drawings in the present application. Figures 1-8 The hearing biomimetic voiceprint recognition method, device, electronic equipment and storage medium of the present application are described.

[0043] Figure 1 is a flow chart of a hearing biomimetic voiceprint recognition method according to an exemplary embodiment. As shown in Figure 1 In an exemplary embodiment, the hearing biomimetic voiceprint recognition method includes steps 110 to 130, which are described in detail as follows.

[0044] Step 110: obtaining a to-be-recognized voice signal and performing cutting processing on the to-be-recognized voice signal to obtain a plurality of to-be-recognized voice segments.

[0045] In the embodiments of the present application, the to-be-recognized voice signal is cut into to-be-recognized voice segments, so that the voiceprint recognition model can more effectively capture local time and frequency features, which are crucial for voiceprint recognition and other tasks.

[0046] In a possible implementation, the to-be-recognized voice signal can be segmented into to-be-recognized voice segments of 200 ms, and in the time dimension, two adjacent to-be-recognized voice segments overlap each other by 10 ms.

[0047] Step 120: inputting each to-be-recognized voice segment into a cochlea biomimetic filter to obtain a voice frequency feature.

[0048] In the embodiments of the present application, the to-be-recognized voice segment is input into the cochlea biomimetic filter, and the cochlea biomimetic filter simulates the selective response of the cochlear basilar membrane to different frequency signals by learning adjustable parameters. Each filter in the cochlea biomimetic filter decomposes the voice signal in a specific frequency range to obtain a decomposed voice frequency component, and each component corresponds to a voice frequency feature in a specific frequency range.

[0049] Step 130: inputting the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result, wherein the voiceprint recognition model is obtained based on a cochlear nucleus biomimetic network.

[0050] The voiceprint recognition model (CNAN, Cochlear Nucleus-inspired Auditory Network) provided in the embodiment of the present application is a bionic computing model based on deep learning, and is used for voiceprint recognition. The voiceprint recognition model is obtained based on a cochlear nucleus bionic network. The cochlear nucleus bionic network can perform multi-level feature extraction on input speech frequency features, capture time sequence features and spatial features of signals, and then obtain a probability distribution of a speaker identity corresponding to the speech signal, so as to realize voiceprint recognition.

[0051] In the embodiment of the present application, the initial feature set related to the original voice signal to be recognized is constructed by dividing the voice signal to be recognized into voice segments to be recognized. The voice segments to be recognized are input into a cochlear bionic filter. The cochlear bionic filter directly extracts speech frequency features from the original speech waveform based on the cochlear basilar membrane characteristics, simulates the response of hair cells to specific frequencies, and efficiently and accurately extracts speech frequency features. Then, the voiceprint recognition model provided in the present application is used for voiceprint recognition. Higher adaptability to noise and complex environments is achieved, so as to improve the reliability and efficiency of voiceprint recognition.

[0052] In an example embodiment of the present application, the voiceprint recognition model comprises a first cochlear nucleus bionic network, a fully connected network and an output layer connected in sequence.

[0053] The first cochlear nucleus bionic network is used for extracting time sequence features and spatial features based on the speech frequency features, obtaining a time sequence feature representation based on the time sequence features and spatial features, and performing weighted fusion processing on the time sequence feature representation and the speech frequency features to obtain a speech feature representation.

[0054] The fully connected network and the output layer are used for sequentially processing the speech feature representation to obtain the voiceprint recognition result.

[0055] In the embodiment of the present application, the structure of the voiceprint recognition model is as shown in Figure 2 The speech frequency features obtained based on the cochlear bionic filter are input into the cochlear nucleus bionic network for multi-level feature extraction, capture of time sequence features and spatial features of signals, and then obtaining a time sequence feature representation of the speech signal. The time sequence feature representation and the speech frequency features are then weighted and fused to obtain a speech feature representation.

[0056] The speech feature representation after weighted fusion is input into the fully connected network for further processing. Finally, the probability distribution of the speaker identity corresponding to the speech signal is calculated through the output layer, and the maximum value is taken to obtain the speaker identity index, that is, the final voiceprint recognition result.

[0057] Specifically, the weighted and fused speech feature representation is flattened into a one-dimensional feature vector, and the full connection network calculates the response score of each speaker based on the one-dimensional feature vector. Based on the response score, the output layer is used to calculate the probability distribution of each speaker, and the final voiceprint recognition result is obtained.

[0058] After the speech signal is integrated in the inferior colliculus, it continues to pass along the auditory pathway to the medial geniculate body and auditory cortex. Therefore, in order to simulate this process, the embodiment of the present application uses a multi-layer full connection network to complete the final classification task of the auditory cortex.

[0059] Specifically, the full connection network calculates as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] wherein, , , , , , , , is a trainable weight matrix and an error term, N represents the total number of classes in voiceprint recognition, is the response score of each speaker.

[0067] In an example embodiment of the present application, the first cochlear nucleus biomimetic network includes a spiral ganglion neuron layer, a stellate cell layer, an attention mechanism, and a bush cell layer, the attention mechanism is connected with the spiral ganglion neuron layer, the stellate cell layer, and the bush cell layer respectively, and the spiral ganglion neuron layer is connected with the stellate cell layer.

[0068] The spiral ganglion neuron layer is configured to perform a convolution operation in a time direction based on the speech frequency feature to obtain a first biomimetic feature.

[0069] The stellate cell layer is configured to perform a convolution operation in a time direction based on the first biomimetic feature to obtain a second biomimetic feature.

[0070] The attention mechanism is used for weighted fusion through attention parameters based on the voice frequency features and the second bionic features to obtain intermediate features.

[0071] The plexiform cell layer is used for performing convolution operation in a time direction based on the intermediate features to obtain the voice feature representation.

[0072] In the embodiment of the present application, the auditory nerve fibers are in the ventral cochlear nucleus, and dominate the main cell types in the core area, i.e., spherical plexiform cells, globular plexiform cells, T star-shaped cells and D star-shaped cells. In the core area of the ventral cochlear nucleus, the D star-shaped cells dominate the plexiform cells and the T star-shaped cells to form a local regulation loop.

[0073] As shown in Figure 3 , Figure 3 The ventral cochlear nucleus local regulation loop provided by the present application is shown in the figure, wherein Sp is a spherical plexiform cell, Ds is a D star-shaped cell, Ts is a T star-shaped cell, I-SG is a spiral ganglion neuron, and IHC is an inner hair cell. The main exogenous excitation input of the ventral cochlear nucleus is the nerve fiber emitted by the spiral ganglion neuron connected with the cochlear inner hair cell, and the neurotransmitter released by the nerve fiber is glutamic acid. The cochlear bionic filter directly extracts the frequency features from the original voice waveform with high efficiency and accuracy based on the cochlear basilar membrane characteristics, and has bionic interpretability. Based on this, the cochlear nucleus bionic network is proposed in the embodiment of the present application, and cochlear nucleus signal processing is added before the next layer processing, so as to increase the interpretability of the model signal processing.

[0074] In a specific embodiment, the connection mode between neurons can be divided into three categories, which are full connection mode, multi-synapse connection projection mode and single-synapse connection mode. In the full connection mode, each star-shaped cell receives synaptic projection of all spiral ganglion neurons, and each plexiform cell receives synaptic projection of all star-shaped cells, which can be used as a transition under the condition that the specific connection mode has not been anatomized in the biological field at present; in the multi-synapse connection projection mode, the connection mode of a small receptive field is realized; in the single-synapse connection mode, each star-shaped cell is projected by a single spiral ganglion neuron, and each plexiform cell also receives projection of a single star-shaped cell. Finally, each neuron in the plexiform cell layer receives projection of a single spiral neuron and performs fusion processing with the projection of the star-shaped cell, as shown in Figure 4 , Figure 4 The three connection modes of the neurons in the cochlear nucleus bionic network provided by the present application are shown in the figure.

[0075] The spiral ganglion neuron layer receives the output of the filter bank , as the input signal of the cochlear nucleus bionic network, continue to pass along the auditory pathway. The spiral ganglion neuron in the spiral ganglion neuron layer projects synaptic connection to the stellate cell and the bushy cell, and the bushy cell is the main uplink information transmission nucleus. The bushy cell receives the inhibitory information from the stellate cell before transmitting the received information to the next layer, and then transmits the information to the next layer. Thus, the information transmission calculation of the cochlear nucleus bionic network is composed. Therefore, based on the above three connection modes, the connection mode between the neurons in the cochlear nucleus bionic network provided by the embodiment of the application adopts a full connection mode, a two synapse connection mode and a single synapse connection mode, as shown in Figure 4 The full connection mode and the prior art use the same transfer function, that is, a one-dimensional convolution is used to summarize the information of all neurons in the previous layer, and a convolution operation is performed in the time dimension. The two synapse connection mode and the single synapse connection mode use a more sparse way to transmit information, that is, the two synapse connection mode processes the information of two neurons in the previous layer, and performs a convolution operation in the time dimension, and the single synapse connection mode only processes the information of one neuron in the previous layer, and performs a convolution operation in the time dimension.

[0076] The calculation formula of the cochlear nucleus bionic network is as follows:

[0077] In the full connection mode (one-dimensional convolution):

[0078] ;

[0079] ;

[0080] In the two synapse connection mode:

[0081] ;

[0082] ;

[0083] In the single synapse connection mode:

[0084] ;

[0085] ;

[0086] wherein, represents a convolution operation in the time direction, represents a column merging operation, represents an activation function, , , , is a trainable multi-channel convolution kernel, is a trainable error term.

[0087] Figure 5 A structural diagram of a voiceprint recognition model provided by the present application is shown in FIG. 1. As shown in FIG. 1, the spiral ganglion neuron layer performs convolution operation in the time direction based on the frequency characteristics of the speech signal to obtain first biomimetic features, and the stellate cell layer performs convolution operation in the time direction based on the first biomimetic features to obtain second biomimetic features, thereby further refining the feature information of the speech signal. Through multi-level convolution operation, the features related to the speech signal are updated and integrated, and this process can dynamically capture the time sequence and frequency neighborhood information of the speech signal, and gradually strengthen the important speech features. Figure 5

[0088] Before the information received by the bush cell is transmitted to the next layer, the bush cell receives inhibition information from the stellate cell, and then transmits the information to the next layer. In the embodiment of the present application, the attention mechanism and the convolution layer are fused to process the time sequence feature representation of the speech signal, and the weighted speech feature representation is obtained, thereby providing accurate local information for voiceprint recognition. The weighted speech feature representation not only reflects the time sequence features and spatial features of different frequency components in the speech signal, but also includes the inter-frequency correlation information updated through the aggregation process. This feature extraction method ensures that the unique features of each speech signal can be fully preserved, and provides multi-level speech function information for subsequent voiceprint recognition, thereby strengthening the perception ability of the model to key features.

[0089] In the embodiment of the present application, the calculation method of the fusion of the attention mechanism and the convolution layer is as follows:

[0090] In the full connection mode:

[0091] ;

[0092] ;

[0093] In the two synapse connection mode:

[0094] ;

[0095] ;

[0096] In the single synapse connection mode:

[0097] ;

[0098] ;

[0099] wherein, , is a trainable attention parameter, is a trainable error term.

[0100] ​In an exemplary embodiment of the present invention, the output layer includes a Softmax function or an additive margin Softmax function.

[0101] In this embodiment of the invention, the additive margin Softmax function is a method to improve inter-class separation by adding a fixed angular margin to the decision boundary, based on the traditional Softmax function. This improvement enables the model to better distinguish feature vectors of different classes in the feature space.

[0102] In an exemplary embodiment of the present invention, the voiceprint recognition model further includes a second cochlear nucleus bionic network, which is connected between the first cochlear nucleus bionic network and the fully connected network.

[0103] In this embodiment of the invention, since nuclei such as the upper olive complex have similar internal neuronal connection patterns, therefore, as Figure 6 As shown, a second cochlear nucleus bionic network, identical to the first cochlear nucleus bionic network, is added between the first cochlear nucleus bionic network and the fully connected network. The output of the first cochlear nucleus bionic network is then processed to extract its temporal feature representation, followed by weighted fusion processing. The output is denoted as... .

[0104] In an exemplary embodiment of the present invention, the cochlear bionic filter includes a filter bank composed of multiple filters with different response frequencies; wherein each filter extracts the speech frequency feature corresponding to the response frequency from the speech segment to be recognized.

[0105] In this embodiment of the invention, the cochlear bionic filter, based on the characteristics of the cochlear basilar membrane, simulates the response of hair cells to specific frequencies, directly and efficiently extracting frequency features from the original speech waveform, exhibiting bionic interpretability. Therefore, the speech segment to be recognized is input into the cochlear bionic filter, and adjustable parameters are learned to simulate the selective response of the cochlear basilar membrane to signals of different frequencies. Each filter decomposes the speech signal within a specific frequency range, obtaining the decomposed speech frequency components, each component corresponding to features within a specific frequency range.

[0106] In one possible implementation, drawing on theory from digital signal processing, a bandpass filter is generated using two low-pass filters. In practice, the bandpass filter function in the time domain is:

[0107] ;

[0108] in, It is a learnable low cutoff frequency and a high cutoff frequency. Indicates time parameter, Given a discrete-time speech signal , and a band-pass filter with a length of The filtered output is:

[0109] ;

[0110] Further, m filters with different response frequencies are constructed, where m is a natural number, to obtain a filter bank The filter bank completes the response of the hair cell to a specific frequency to obtain an output , is the length of the segmented frame-level audio data.

[0111] In an example embodiment of the present application, the training process of the voiceprint recognition model includes the following steps, which are described in detail as follows.

[0112] Obtain a training set; wherein the training set includes a plurality of training samples.

[0113] Based on the training samples, a loss function is constructed.

[0114] The training samples in the training set are input into a basic voiceprint recognition model, and the loss function is minimized as the training target for iterative training until the training is completed, to obtain the voiceprint recognition model; wherein the basic voiceprint recognition model is obtained based on a cochlear nucleus bionic network.

[0115] In the embodiment of the present application, the loss function of the voiceprint recognition model is set, and the loss function is represented by the following formula:

[0116] ;

[0117] wherein, represents the total number of training samples in the training set, represents the total number of classes in the voiceprint recognition, represents the real probability distribution, that is, if the voice signal is the speaking voice of the first person, then .

[0118] The basic voiceprint recognition model is trained with the training target of minimizing the loss function, and the weight matrix of the basic voiceprint recognition model is iteratively updated to obtain the voiceprint recognition model.

[0119] ​The hearing biomimetic voiceprint recognition device provided by the present application is described below, and the hearing biomimetic voiceprint recognition device described below can be correspondingly referred to the hearing biomimetic voiceprint recognition method described above. It should be noted that the device provided in the following embodiment and the method provided in the above embodiment belong to the same concept, and the specific manner in which each module and unit performs operations has been described in detail in the method embodiment, which will not be described here.

[0120] In an example embodiment of the present application, please refer to Figure 7 , Figure 7 is a hearing biomimetic voiceprint recognition device according to an example embodiment, comprising:

[0121] The first acquisition module 710 is configured to acquire a to-be-identified voice signal, and perform cutting processing on the to-be-identified voice signal to obtain a plurality of to-be-identified voice segments;

[0122] The first input module 720 is configured to input each of the to-be-identified voice segments into a cochlea biomimetic filter to obtain a voice frequency feature;

[0123] The second input module 730 is configured to input the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result; wherein the voiceprint recognition model is obtained based on a cochlear nucleus biomimetic network.

[0124] In an example embodiment of the present application, the voiceprint recognition model comprises a first cochlear nucleus biomimetic network, a fully connected network and an output layer connected in sequence;

[0125] The first cochlear nucleus biomimetic network is configured to extract a time sequence feature and a spatial feature based on the voice frequency feature, obtain a time sequence feature representation based on the time sequence feature and the spatial feature, and perform weighted fusion processing on the time sequence feature representation and the voice frequency feature to obtain a voice feature representation;

[0126] The fully connected network and the output layer are configured to sequentially process the voice feature representation to obtain the voiceprint recognition result.

[0127] In an example embodiment of the present application, the first cochlear nucleus biomimetic network comprises a spiral ganglion neuron layer, a stellate cell layer, an attention mechanism and a cluster cell layer, the attention mechanism is connected with the spiral ganglion neuron layer, the stellate cell layer and the cluster cell layer respectively, and the spiral ganglion neuron layer is connected with the stellate cell layer;

[0128] The spiral ganglion neuron layer is configured to perform convolution operation in the time direction based on the voice frequency feature to obtain a first biomimetic feature;

[0129] The star cell layer is configured to perform a convolution operation in a time direction based on the first biomimetic feature to obtain a second biomimetic feature.

[0130] The attention mechanism is configured to perform weighted fusion through attention parameters based on the voice frequency feature and the second biomimetic feature to obtain an intermediate feature.

[0131] The cluster cell layer is configured to perform a convolution operation in a time direction based on the intermediate feature to obtain the voice feature representation.

[0132] In an example embodiment of the present application, the output layer includes a Softmax function or an additive margin Softmax function.

[0133] In an example embodiment of the present application, the voiceprint recognition model further includes a second cochlear nucleus biomimetic network connected between the first cochlear nucleus biomimetic network and the fully connected network.

[0134] In an example embodiment of the present application, the cochlear biomimetic filter includes a filter bank composed of a plurality of filters with different response frequencies; wherein each filter extracts the voice frequency feature of the corresponding response frequency from the to-be-identified voice segment.

[0135] In an example embodiment of the present application, the auditory biomimetic voiceprint recognition device further includes:

[0136] The second acquisition module is configured to acquire a training set; wherein the training set includes a plurality of training samples.

[0137] The framework module is configured to construct a loss function based on the training samples.

[0138] The iterative training module is configured to input the training samples in the training set into a basic voiceprint recognition model to perform iterative training with the minimization of the loss function as a training target until the training is completed to obtain the voiceprint recognition model; wherein the basic voiceprint recognition model is obtained based on a cochlear nucleus biomimetic network.

[0139] Figure 8 An example of an entity structure diagram of an electronic device is shown in FIG. 1. Figure 8As shown, the electronic device can include a processor 810, a communications interface 820, a memory 830, and a communications bus 840, wherein the processor 810, the communications interface 820, and the memory 830 complete mutual communication through the communications bus 840. The processor 810 can invoke a logical instruction in the memory 830 to execute an auditory biomimetic voiceprint recognition method, which includes: acquiring a to-be-recognized voice signal, and performing cutting processing on the to-be-recognized voice signal to obtain a plurality of to-be-recognized voice segments;

[0140] Inputting each to-be-recognized voice segment into a cochlea biomimetic filter to obtain a voice frequency feature;

[0141] Inputting the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result; wherein the voiceprint recognition model is obtained based on a cochlear nucleus biomimetic network.

[0142] In addition, the logical instruction in the memory 830 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0143] On the other hand, the present application also provides a computer program product, which includes a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program is executed by a processor, the computer can execute the auditory biomimetic voiceprint recognition method provided by the above-mentioned method, which includes: acquiring a to-be-recognized voice signal, and performing cutting processing on the to-be-recognized voice signal to obtain a plurality of to-be-recognized voice segments;

[0144] Inputting each to-be-recognized voice segment into a cochlea biomimetic filter to obtain a voice frequency feature;

[0145] Inputting the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result; wherein the voiceprint recognition model is obtained based on a cochlear nucleus biomimetic network.

[0146] In yet another aspect, the present application also provides a non-transitory computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the method for recognizing an auditory biomimetic voice provided by the above-mentioned methods, the method comprising: obtaining a to-be-recognized voice signal, and performing cutting processing on the to-be-recognized voice signal to obtain a plurality of to-be-recognized voice segments;

[0147] inputting each of the to-be-recognized voice segments into a cochlea biomimetic filter to obtain a voice frequency feature;

[0148] inputting the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result, wherein the voiceprint recognition model is obtained based on a cochlear nucleus biomimetic network.

[0149] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0150] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus a necessary general hardware platform, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in terms of the contribution to the prior art, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0151] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An aural biomimetic voiceprint recognition method, characterized in that, The method comprises: acquiring a to-be-recognized voice signal, and performing cutting processing on the to-be-recognized voice signal to obtain a plurality of to-be-recognized voice segments; inputting each to-be-recognized voice segment into a cochlea bionic filter to obtain a voice frequency feature; inputting the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result; wherein the voiceprint recognition model is obtained based on a cochlear nucleus bionic network; the voiceprint recognition model comprises a first cochlear nucleus bionic network, a fully connected network and an output layer connected in sequence; the first cochlear nucleus bionic network is configured to extract a time sequence feature and a spatial feature based on the voice frequency feature, obtain a time sequence feature representation based on the time sequence feature and the spatial feature, and perform weighted fusion processing on the time sequence feature representation and the voice frequency feature to obtain a voice feature representation; the fully connected network and the output layer are configured to sequentially process the voice feature representation to obtain the voiceprint recognition result; the first cochlear nucleus bionic network comprises a spiral ganglion neuron layer, a stellate cell layer, an attention mechanism and a cluster cell layer, the attention mechanism is connected with the spiral ganglion neuron layer, the stellate cell layer and the cluster cell layer respectively, and the spiral ganglion neuron layer is connected with the stellate cell layer; the spiral ganglion neuron layer is configured to perform convolution operation in a time direction based on the voice frequency feature to obtain a first bionic feature; the stellate cell layer is configured to perform convolution operation in a time direction based on the first bionic feature to obtain a second bionic feature; the attention mechanism is configured to perform weighted fusion on the voice frequency feature and the second bionic feature through an attention parameter to obtain an intermediate feature; the cluster cell layer is configured to perform convolution operation in a time direction based on the intermediate feature to obtain the voice feature representation.

2. The aural biomimetic voiceprint recognition method of claim 1, wherein, The output layer comprises a Softmax function or an additive margin Softmax function.

3. The aural biomimetic voiceprint recognition method of claim 1, wherein, The voiceprint recognition model further comprises a second cochlear nucleus bionic network connected between the first cochlear nucleus bionic network and the fully connected network.

4. The aural biomimetic voiceprint recognition method of claim 1, wherein, The cochlea bionic filter comprises a filter bank composed of a plurality of filters with different response frequencies; wherein each filter extracts the voice frequency feature of the corresponding response frequency from the to-be-recognized voice segment.

5. The aurally biomimetic voiceprint recognition method of any one of claims 1 to 4, characterized in that, The training process of the voiceprint recognition model comprises: acquiring a training set; wherein the training set comprises a plurality of training samples; constructing a loss function based on the training samples; inputting the training samples in the training set into a basic voiceprint recognition model to perform iterative training with the minimization of the loss function as a training target until the training is completed to obtain the voiceprint recognition model; wherein the basic voiceprint recognition model is obtained based on a cochlear nucleus bionic network.

6. An aural biomimetic voiceprint recognition device, characterized by, The method comprises: a first acquisition module configured to acquire a to-be-recognized voice signal, and perform cutting processing on the to-be-recognized voice signal to obtain a plurality of to-be-recognized voice segments; a first input module configured to input each to-be-recognized voice segment into a cochlea bionic filter to obtain a voice frequency feature; The second input module is configured to input the voice frequency feature into a pre-trained voiceprint recognition model to obtain a voiceprint recognition result, wherein the voiceprint recognition model is obtained based on a cochlear nuclear bionic network. The voiceprint recognition model comprises a first cochlear nuclear bionic network, a fully connected network and an output layer connected in sequence. The first cochlear nuclear bionic network is configured to extract time sequence features and spatial features based on the voice frequency feature, obtain time sequence feature representation based on the time sequence features and the spatial features, and perform weighted fusion processing on the time sequence feature representation and the voice frequency feature to obtain voice feature representation. The fully connected network and the output layer are configured to sequentially process the voice feature representation to obtain the voiceprint recognition result. The first cochlear nuclear bionic network comprises a spiral ganglion neuron layer, a stellate cell layer, an attention mechanism and a cluster cell layer, the attention mechanism is connected with the spiral ganglion neuron layer, the stellate cell layer and the cluster cell layer respectively, and the spiral ganglion neuron layer is connected with the stellate cell layer. The spiral ganglion neuron layer is configured to perform convolution operation in the time direction based on the voice frequency feature to obtain first bionic features. The stellate cell layer is configured to perform convolution operation in the time direction based on the first bionic features to obtain second bionic features. The attention mechanism is configured to perform weighted fusion on the voice frequency feature and the second bionic features through attention parameters to obtain intermediate features. The cluster cell layer is configured to perform convolution operation in the time direction based on the intermediate features to obtain the voice feature representation.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the hearing bionic voiceprint recognition method according to any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the hearing bionic voiceprint recognition method according to any one of claims 1 to 5.

Images