A voice detection method and apparatus

Abstract

This invention provides a voice detection method and apparatus, relating to the field of sound detection technology. It includes: inputting a speech signal to be detected into an interpretable convolutional filter bank to obtain a multi-band speech signal; extracting the temporal and time-frequency features of the multi-band speech signal using a first path and a second path of a one-dimensional deep separable convolutional neural network, respectively, and combining them to obtain a multi-band speech signal feature vector; classifying the multi-band speech signal feature vector using a fully connected network to obtain a classification result, and using the classification result as an anomaly index of the speech signal to be detected. This invention uses an interpretable convolutional filter bank to better capture the frequency band information of the speech signal to be detected, while the one-dimensional deep separable convolutional neural network can extract the temporal and time-frequency features of the multi-band speech signal separately, increasing the effectiveness of feature extraction, making the method more interpretable, and improving the reliability of the voice detection results.

Description

Technical Field

[0001] This invention relates to the field of sound detection technology, and in particular to a method and apparatus for voice detection. Background Technology

[0002] Voice is an essential tool for communication, and automated voice detection is a crucial method for assessing voice quality. In recent decades, it has gained increasing attention due to its non-invasiveness, objectivity, and convenience. This technology collects the voice of a test subject through a microphone and analyzes the sound to determine if the subject has voice problems.

[0003] Currently, this technology mainly employs two methods. One is a feature-based approach, which first extracts different features based on the characteristics of the sound, and then uses these features to train a machine learning classifier for sound classification. The features used in this method include perturbation features (fundamental frequency perturbation, amplitude perturbation, signal-to-noise ratio and harmonic-to-noise ratio), spectral / cepteral features (spectral tilt, cepstral peak prominence, harmonic amplitude, etc.), glottal features (opening quotient, amplitude quotient, closing quotient, velocity quotient, etc.), nonlinear features (maximum Lyapunov exponent, correlation dimension, nonlinear recursive quantization analysis features, etc.), and multi-band features (Mel frequency cepstral coefficients, gammatone cepstral coefficients, and gammatone spectral height). Because the features used in the feature-based approach are proposed based on the differences in different voice problems, they are usually interpretable. However, this method requires the selection of a suitable classifier; the same features may yield significantly different results under different classifiers, meaning the reliability of voice detection results is not high. The other method is a neural network-based approach, which directly inputs the sound into the neural network, allowing the network to extract deep features and perform classification. However, while two-dimensional neural networks are primarily used for applications such as object detection in two-dimensional images, applying them to voice detection would miss the essential nature of speech as a one-dimensional time series. One-dimensional neural networks are better suited to time series processing than two-dimensional neural networks, but simply stacking convolutional neural networks or deep neural networks is ineffective for voice detection. Because neural networks automatically acquire appropriate features from the input signal and then use fully connected layers to make judgments, their effectiveness is often higher than feature-based methods. However, the automatic feature acquisition by neural networks lacks interpretability and is more prone to overfitting to training data.

[0004] In summary, existing voice detection methods suffer from low feature extraction effectiveness, lack of interpretability, and low reliability of voice detection results. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problems of low feature extraction effectiveness, lack of interpretability and low reliability of voice detection results in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides a voice detection method, comprising:

[0007] The speech signal to be detected is input into an interpretable convolutional filter bank to obtain a multi-band speech signal;

[0008] The time features of the multi-band speech signal are extracted using the first path of the one-dimensional deep separable convolutional neural network, and the time-frequency features of the multi-band speech signal are extracted using the second path of the one-dimensional deep separable convolutional neural network. The time features of the multi-band speech signal and the time-frequency features of the multi-band speech signal are combined as the feature vector of the multi-band speech signal.

[0009] The feature vectors of the multi-band speech signal are classified using a fully connected network to obtain a classification result, and the classification result is used as the anomaly index of the speech signal to be detected.

[0010] In one embodiment of the present invention, before inputting the speech signal to be detected into an interpretable convolutional filter bank to obtain a multi-band speech signal, the method further includes: assigning different weights to the filters in the interpretable convolutional filter bank.

[0011] In one embodiment of the present invention, the step of inputting the speech signal to be detected into an interpretable convolutional filter bank to obtain a multi-band speech signal includes:

[0012] The speech signal to be detected is divided into M frames;

[0013] The M frames of speech signals to be detected are input into the interpretable convolutional filter bank to obtain M groups of multi-frequency band speech signals.

[0014] In one embodiment of the present invention, the step of extracting the temporal features of the multi-band speech signal using a first path of a one-dimensional deep separable convolutional neural network, extracting the time-frequency features of the multi-band speech signal using a second path of the one-dimensional deep separable convolutional neural network, and combining the temporal features of the multi-band speech signal with the time-frequency features of the multi-band speech signal as a feature vector includes:

[0015] The M groups of multi-frequency band speech signals are input into the first path of the one-dimensional deep separable convolutional neural network to extract the time features of the multi-frequency band speech signals, thereby obtaining the time feature vector of the M groups of multi-frequency band speech signals.

[0016] The M groups of multi-frequency band speech signals are input into the second path of the one-dimensional deep separable convolutional neural network to extract the time-frequency features of the multi-frequency band speech signals, thereby obtaining the time-frequency feature vector of the M groups of multi-frequency band speech signals.

[0017] The time feature vector and time-frequency feature vector of the M groups of multi-frequency band speech signals are combined to obtain the M groups of multi-frequency band speech signal feature vectors;

[0018] The first path includes a first separable convolution module, a second separable convolution module, and a third separable convolution module, while the second path includes a fourth separable convolution module, a first depthwise separable convolution module, and a second depthwise separable convolution module.

[0019] In one embodiment of the present invention, the step of inputting the M groups of multi-frequency band speech signals into the first path of the one-dimensional deep separable convolutional neural network to extract the time features of the multi-frequency band speech signals, and obtaining the time feature vector of the M groups of multi-frequency band speech signals, includes:

[0020] The multi-band speech signal is input into the first separable convolution module to obtain the first multi-band speech signal feature vector, and the average value and variance of the first multi-band speech signal feature vector are calculated.

[0021] The first multi-band speech signal feature vector is input into the second separable convolution module to obtain the second multi-band speech signal feature vector.

[0022] The second multi-band speech signal feature vector is input into the third separable convolution module to obtain the third multi-band speech signal feature vector.

[0023] The average value and variance of the first multi-band speech signal feature vector are combined with the third multi-band speech signal feature vector to form the time feature vector of the multi-band speech signal.

[0024] In one embodiment of the present invention, the step of inputting the M groups of multi-frequency band speech signals into the second path of the one-dimensional deep separable convolutional neural network to extract the time-frequency features of the multi-frequency band speech signals, and obtaining the time-frequency feature vectors of the M groups of multi-frequency band speech signals, includes:

[0025] The multi-band speech signal is input into the first depthwise separable convolution module to obtain the fourth multi-band speech signal feature vector, and the mean and variance of the fourth multi-band speech signal feature vector are calculated.

[0026] The fourth multi-band speech signal feature vector is input into the fourth separable convolution module to obtain the fifth multi-band speech signal feature vector.

[0027] The fifth multi-band speech signal feature vector is input into the second depthwise separable convolution module to obtain the sixth multi-band speech signal feature vector.

[0028] The average value and variance of the fourth multi-band speech signal feature vector are combined with the sixth multi-band speech signal feature vector to form the time-frequency feature vector of the multi-band speech signal.

[0029] In one embodiment of the present invention, the step of classifying the feature vector of the multi-band speech signal using a fully connected network to obtain a classification result, and using the classification result as an anomaly index of the speech signal to be detected, includes:

[0030] The M groups of multi-band speech signal feature vectors are input into a fully connected network for classification to obtain M groups of classification results.

[0031] The average of the M classification results is taken as the anomaly index of the speech signal to be detected.

[0032] In one embodiment of the present invention, after classifying the feature vector of the multi-band speech signal using a fully connected network to obtain a classification result and using the classification result as the anomaly index of the speech signal to be detected, the method further includes: comparing the anomaly index of the speech signal to be detected with a preset threshold to determine the state of the speech signal to be detected.

[0033] The present invention also provides a voice detection device, comprising:

[0034] The filtering module is used to input the speech signal to be detected into an interpretable convolutional filter bank to obtain a multi-band speech signal;

[0035] The feature extraction module is used to extract the time features of the multi-band speech signal using the first path of the one-dimensional deep separable convolutional neural network, extract the time-frequency features of the multi-band speech signal using the second path of the one-dimensional deep separable convolutional neural network, and combine the time features of the multi-band speech signal with the time-frequency features of the multi-band speech signal as a feature vector of the multi-band speech signal.

[0036] The classification module is used to classify the feature vector of the multi-frequency band speech signal using a fully connected network to obtain a classification result, and uses the classification result as the anomaly index of the speech signal to be detected.

[0037] In one embodiment of the present invention, a comparison module is further included, which is used to compare the anomaly index of the voice signal to be detected with a preset threshold to determine the state of the voice signal to be detected.

[0038] The voice detection method of this invention inputs the speech signal to be detected into an interpretable convolutional filter bank to obtain a multi-band speech signal. The first path of a one-dimensional deep separable convolutional neural network (DSN) is used to extract the temporal features of the multi-band speech signal, and the second path of the same DSN is used to extract the time-frequency features. The temporal and time-frequency features are combined to form the multi-band speech signal feature vector. Finally, a fully connected network is used to classify the feature vector to obtain a classification result, which is then used as the anomaly index of the speech signal to be detected. Because the interpretable convolutional filter bank is more effective at capturing frequency band information and is more interpretable than traditional convolutional neural networks in speech signal processing, and because the one-dimensional deep separable convolutional neural network includes a first path and a second path, which can be used to extract the temporal and time-frequency features of the speech signal to be detected, the effectiveness and interpretability of feature extraction are improved, making the voice detection results more reliable. Attached Figure Description

[0039] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein...

[0040] Figure 1 This is a flowchart of the voice detection method provided by the present invention;

[0041] Figure 2 This is a schematic diagram of the interpretable convolutional filter bank (SincNet) structure;

[0042] Figure 3 This is a schematic diagram of the convolutional block structure of a one-dimensional convolutional neural network;

[0043] Figure 4 This is a schematic diagram of the convolutional block structure of a one-dimensional depthwise separable convolutional network.

[0044] Figure 5 This is a schematic diagram of a one-dimensional depthwise separable convolutional neural network structure.

[0045] Figure 6 This is a schematic diagram of a separable convolutional module structure;

[0046] Figure 7 This is a schematic diagram of a depth-separable convolutional module structure;

[0047] Figure 8 This is a schematic diagram of the amplitude-frequency response of SincNet;

[0048] Figure 9 A schematic diagram of the amplitude-frequency response of an interpretable convolutional filter bank (AT-SincNet) provided by the present invention;

[0049] Figure 10 This is a diagram comparing the training losses of AT-SincNet and SincNet.

[0050] Figure 11 This is a diagram showing the comparison of test losses between AT-SincNet and SincNet.

[0051] Figure 12 This is a schematic diagram illustrating the principle of another voice detection method provided by the present invention;

[0052] Figure 13 A schematic diagram illustrating the importance of features extracted by the voice detection method provided by this invention;

[0053] Figure 14 This is a schematic diagram of the voice detection device provided by the present invention. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0055] Example 1:

[0056] Reference Figure 1 As shown, the voice detection method provided by the present invention includes:

[0057] S10: Input the speech signal to be detected into an interpretable convolutional filter bank to obtain a multi-band speech signal.

[0058] Interpretable convolutional filter banks (SincNet) were first proposed in 2018, and their structure is as follows: Figure 2 As shown, this filter bank uses a series of predefined functions g to perform convolution operations, and its calculation formula is as follows:

[0059] y(n)=x(n)*[g(n,θ)·w(n)],

[0060] Where x(n) and y(n) represent the input and output of the filter bank, respectively, and w(n) is the Hamming window, which allows the finite-length filter to approximate the performance of an ideal filter of infinite length. Unlike traditional filters, g(n,θ) depends only on a few learnable parameters θ. If g is defined using a rectangular filter bank with learnable low cutoff frequencies f1 and f2, its time-domain and frequency-domain expressions are as follows:

[0061] g(n,f1,f2)=2f2sinc(2πf2n)-2f1sinc(2πf1n),

[0062]

[0063] Where sinc(x) = sin(x) / x, and rect(·) represents the rectangular function.

[0064] Compared to traditional convolutional neural networks (CNNs), SincNet has the advantages of faster convergence, fewer network parameters, and higher computational efficiency. In addition, SincNet is more likely to capture narrowband information in speech signal processing, while CNNs tend to focus on the low-frequency part of speech information. Therefore, this embodiment uses SincNet to make the voice detection results more interpretable.

[0065] S12: Extract the time features of the multi-band speech signal using the first path of a one-dimensional deep separable convolutional neural network, extract the time-frequency features of the multi-band speech signal using the second path of the one-dimensional deep separable convolutional neural network, and combine the time features and time-frequency features of the multi-band speech signal as the feature vector of the multi-band speech signal.

[0066] The convolutional blocks of a classic one-dimensional convolutional neural network are as follows: Figure 3 As shown, the output of its convolutional block can be expressed as:

[0067] O k =∑ n conv(I n ,F k ),

[0068] In this method, a multi-channel time signal is convolved and weighted through multiple filters to obtain a multi-channel signal output. This method can make the most of the information from each frequency band. However, the combination and segmentation of multiple frequency bands prevents this convolutional network from obtaining specific outputs. The output of each channel has no corresponding relationship with the frequency band, which reduces the interpretability of the neural network.

[0069] Based on this, the embodiments of this application employ a one-dimensional depthwise separable convolutional network to extract features of multi-band speech signals, and its convolutional blocks are as follows: Figure 4 As shown, this includes depthwise convolution and pointwise convolution. The output of each convolutional block can be expressed as:

[0070] O n =conv(I n ,F n ),

[0071] O k =∑ n W k,n *O n +b k .

[0072] Depthwise convolution is performed on a channel-by-channel basis, so information from different channels will not interfere with each other. If it is necessary to fuse information from different channels, pointwise convolution can be used to combine the convolution results of different channels.

[0073] like Figure 5 The diagram shows the structure of a one-dimensional depthwise separable convolutional neural network provided in an embodiment of this application, including a first path and a second path.

[0074] The first path is used to extract the time features of the multi-band speech signal, and the second path is used to extract the time-frequency features of the multi-band speech signal.

[0075] Optionally, the first path includes a first separable convolutional module, a second separable convolutional module, and a third separable convolutional module. The second path includes a first depthwise separable convolutional module, a second depthwise separable convolutional module, and a fourth separable convolutional module.

[0076] Among them, such as Figure 6 As shown, the separable convolutional module includes a one-dimensional separable convolutional layer, a batch normalization layer, and an activation function layer. Figure 7 As shown, the depthwise separable convolution module includes a one-dimensional separable convolutional layer, a one-dimensional pointwise convolutional layer, a batch normalization layer, and an activation function layer.

[0077] For example, based on Figure 5 The one-dimensional depthwise separable convolutional neural network structure shown is implemented as follows:

[0078] S121: Input the multi-band speech signal into the first separable convolution module to obtain the first multi-band speech signal feature vector, and calculate the average value and variance of the first multi-band speech signal feature vector.

[0079] S122: Input the first multi-band speech signal feature vector into the second separable convolution module to obtain the second multi-band speech signal feature vector.

[0080] S123: Input the second multi-band speech signal feature vector into the third separable convolution module to obtain the third multi-band speech signal feature vector.

[0081] S124: Combine the mean and variance of the first multi-band speech signal feature vector with the third multi-band speech signal feature vector to form the time feature vector of the multi-band speech signal.

[0082] The first multi-band speech signal feature vector represents the detailed features of each frequency band speech signal. Considering the dimension of the output feature vector, the mean and variance of the first multi-band speech signal feature vector are used as part of the time feature vector of the multi-band speech signal. The third multi-band speech signal feature vector represents the global features of each frequency band speech signal, reflecting the characteristics of the multi-band speech signal over a longer period of time.

[0083] S125: Input the multi-band speech signal into the first depthwise separable convolution module to obtain the feature vector of the fourth multi-band speech signal, and calculate the mean and variance of the feature vector of the fourth multi-band speech signal;

[0084] S126: Input the fourth multi-band speech signal feature vector into the fourth separable convolution module to obtain the fifth multi-band speech signal feature vector;

[0085] S127: Input the fifth multi-band speech signal feature vector into the second depthwise separable convolution module to obtain the sixth multi-band speech signal feature vector;

[0086] S128: Combine the mean and variance of the fourth multi-band speech signal feature vector with the sixth multi-band speech signal feature vector to form the time-frequency feature vector of the multi-band speech signal.

[0087] S129: Combine the time features and time-frequency features of the multi-band speech signal as the feature vector of the multi-band speech signal.

[0088] The second path is structurally similar to the first path. The main difference is that the second path uses a point convolution of an output channel to determine the relationship between different frequency bands before outputting short-term and long-term time-frequency features.

[0089] Optionally, in other embodiments of this application, the number of layers in the one-dimensional deep separable convolutional neural network can be 1, 2, or others. When using a neural network to extract features, the features extracted from shallow layers are often more detailed. As the number of network layers increases, the features extracted by the network tend to be more inclined towards large-area features rather than detailed features. Therefore, the deep separable convolutional neural network in this embodiment adopts three convolutional layers.

[0090] S13: Use a fully connected network to classify the feature vectors of multi-band speech signals to obtain classification results, and use the classification results as an anomaly index of the speech signal to be detected.

[0091] In some embodiments, step S13 is specifically implemented as follows: inputting the feature vector of the multi-band speech signal into a fully connected network, using the SoftMax function to obtain the classification result based on the nonlinear combination relationship between the feature vectors, and calculating the predicted probability that the feature vector of the multi-band speech signal is classified into each category.

[0092] The formula for calculating the prediction probability is as follows:

[0093]

[0094] Where exp(·) is an exponential function.

[0095] In some embodiments, after step S13, the method further includes: comparing the abnormality index of the voice signal to be detected with a preset threshold to determine the health status of the voice signal to be detected.

[0096] In this embodiment, an interpretable convolutional filter bank can better capture the frequency band information in the speech signal. The first and second paths of a one-dimensional deep separable convolutional neural network are used to extract short-time and long-time temporal and time-frequency features of the multi-band speech signal as feature vectors, improving the effectiveness and interpretability of feature extraction. Finally, a fully connected network is used to classify the multi-band speech signal to obtain the anomaly index of the speech signal to be detected. Therefore, using the interpretable convolutional filter bank and one-dimensional deep separable convolutional neural network provided in this embodiment for voice detection not only improves the effectiveness of feature extraction but also makes it more interpretable, increasing the reliability of the voice detection results.

[0097] Example 2:

[0098] Based on Embodiment 1 above, this application also proposes an improved interpretable convolutional filter bank (AT-SincNet).

[0099] For some complex deep models, the parameter optimization of the existing interpretable convolutional filter bank (SincNet) is slow and limited. In the process of voice detection, its optimization range is only in the tens of hertz, which is not significant compared with the actual frequency range of speech.

[0100] Therefore, the AT-SincNet provided in this embodiment adds a trainable parameter A, i.e., the amplitude of each filter, to SincNet so that when the filter bank divides the speech signal to be detected into frequency bands, it can increase or decrease the amplitude response of each frequency band according to the importance of each frequency band signal, thereby reducing the effect of unimportant frequency bands.

[0101] Specifically, the time-domain and frequency-domain expressions of AT-SincNet are as follows:

[0102] g AT (n,f1,f2)=A[2f2sinc(2πf2n)-2f1sinc(2πf1n)],

[0103]

[0104] like Figure 8 , Figure 9 The figures shown are schematic diagrams of the amplitude-frequency response of SincNet and AT-SincNet, respectively. Figure 10 The diagram shows a comparison of the training losses of SincNet and AT-SincNet. Figure 11 The diagram shows a comparison of the test losses between SincNet and AT-SincNet.

[0105] It can be seen that AT-SincNet has a better fit than SincNet. Due to the greater depth of the network, neither of them has significant optimization for frequency band location and width. However, AT-SincNet can train the amplitude of each filter, thus better reflecting the importance of different frequency bands.

[0106] Example 3:

[0107] Based on Embodiments 1 and 2 above, this embodiment provides another speech detection method, the principle of which is illustrated in the following diagram. Figure 12 As shown, it includes:

[0108] Step 1: Divide the speech signal to be detected into M frames.

[0109] Step 2: Input the M frames of speech signals to be detected into AT-SincNet to obtain M groups of multi-band speech signals.

[0110] Step 3: Input the M groups of multi-frequency band speech signals into the first path of a one-dimensional deep separable convolutional neural network to extract the time features of the multi-frequency band speech signals, and obtain the time feature vectors of the M groups of multi-frequency band speech signals.

[0111] Step 4: Input the M groups of multi-frequency band speech signals into the second path of a one-dimensional deep separable convolutional neural network to extract the time-frequency features of the multi-frequency band speech signals, and obtain the time-frequency feature vectors of the M groups of multi-frequency band speech signals.

[0112] Step 5: Combine the time feature vectors and time-frequency feature vectors of the M groups of multi-frequency band speech signals to obtain the feature vectors of the M groups of multi-frequency band speech signals.

[0113] Step 6: Input the feature vectors of M groups of multi-frequency band speech signals into a fully connected network for classification to obtain the classification results of M groups.

[0114] Step 7: Take the mean of the classification results of M groups as the anomaly index of the speech signal to be detected.

[0115] Step 8: Compare the abnormality index of the voice signal to be detected with a preset threshold to determine the health status of the voice signal to be detected.

[0116] For example, in some embodiments, after obtaining M groups of classification results, each group of classification results can be judged, and then the expected value of the judgment result can be used as the anomaly index of the speech signal to be detected. The anomaly index is compared with a preset threshold p, and the following judgment is made:

[0117] If index > p,

[0118] tjen result = pathological;

[0119] else result = healthy.

[0120] Because the voice is a complex nonlinear signal, even a healthy voice may exhibit abnormal values ​​in certain areas, while a pathological voice may show a normal waveform for a period of time. In some voice detection methods, to reduce the amount of input data, a specific segment of the speech to be detected is chosen for testing. However, due to the characteristics of the voice, this method often leads to erroneous voice detection results.

[0121] Therefore, the voice detection method provided in this embodiment divides a speech signal into several frames, judges each frame separately, and then takes the average of the judgment results of several frames as the abnormality index of the speech signal to be detected. This reduces the impact of some abnormal frames on the voice detection results and improves the reliability of the voice detection results.

[0122] Example 4:

[0123] Based on the above embodiments 1, 2 and 3, this application embodiment also uses the method to conduct voice detection experiments in three voice databases, MEEI, SVD, and HUPA, as well as voice data collected by the First Affiliated Hospital of Soochow University.

[0124] like Figure 13 The diagram illustrates the importance of feature extraction when using this method for voice detection.

[0125] Table 1 shows a comparison of the detection results using this method and other methods in three databases.

[0126] Table 1

[0127]

[0128]

[0129] Table 2 shows a comparison of the results of voice data collected by the First Affiliated Hospital of Soochow University using this method and other methods.

[0130] Table 2

[0131]

[0132] It can be seen that the voice detection method proposed in this application has higher accuracy than other methods when performing voice detection.

[0133] This application also provides a voice detection device, such as Figure 14 As shown, it includes:

[0134] The filtering module 10 is used to input the speech signal to be detected into an interpretable convolutional filter bank to obtain a multi-band speech signal.

[0135] The feature extraction module 20 is used to extract the time features of the multi-band speech signal using the first path of the one-dimensional deep separable convolutional neural network, extract the time-frequency features of the multi-band speech signal using the second path of the one-dimensional deep separable convolutional neural network, and combine the time features and time-frequency features of the multi-band speech signal as the feature vector of the multi-band speech signal.

[0136] The classification module 30 is used to classify the feature vectors of multi-frequency band speech signals using a fully connected network to obtain classification results, and uses the classification results as an anomaly index of the speech signal to be detected.

[0137] Optionally, the voice detection device further includes a comparison module for comparing the abnormality index of the voice signal to be detected with a preset threshold to determine the state of the voice signal to be detected.

[0138] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0140] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0141] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0142] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for detecting voice, characterized in that, include: The speech signal to be detected is input into an interpretable convolutional filter bank to obtain a multi-band speech signal; Specifically, it includes: The speech signal to be detected is divided into M frames; M frames of speech signals to be detected are respectively input into the interpretable convolutional filter group to obtain M groups of multi-frequency band speech signals; The time features of the multi-band speech signal are extracted using the first path of the one-dimensional deep separable convolutional neural network, and the time-frequency features of the multi-band speech signal are extracted using the second path of the one-dimensional deep separable convolutional neural network. The time features of the multi-band speech signal and the time-frequency features of the multi-band speech signal are combined as the feature vector of the multi-band speech signal. The feature vectors of the multi-band speech signal are classified using a fully connected network to obtain a classification result, and the classification result is used as the anomaly index of the speech signal to be detected; specifically, this includes: The M groups of multi-band speech signal feature vectors are input into a fully connected network for classification to obtain M groups of classification results. The average of the M classification results is taken as the anomaly index of the speech signal to be detected.

2. The voice detection method according to claim 1, characterized in that, Before inputting the speech signal to be detected into an interpretable convolutional filter bank to obtain a multi-band speech signal, the method further includes: assigning different weights to the filters in the interpretable convolutional filter bank.

3. The voice detection method according to claim 1, characterized in that, The step of extracting the temporal features of the multi-band speech signal using a first path of a one-dimensional deep separable convolutional neural network, extracting the time-frequency features of the multi-band speech signal using a second path of the same network, and combining the temporal features and time-frequency features of the multi-band speech signal as a feature vector includes: The M groups of multi-frequency band speech signals are input into the first path of the one-dimensional deep separable convolutional neural network to extract the time features of the multi-frequency band speech signals, thereby obtaining the time feature vector of the M groups of multi-frequency band speech signals. The M groups of multi-frequency band speech signals are input into the second path of the one-dimensional deep separable convolutional neural network to extract the time-frequency features of the multi-frequency band speech signals, thereby obtaining the time-frequency feature vector of the M groups of multi-frequency band speech signals. The time feature vector and time-frequency feature vector of the M groups of multi-frequency band speech signals are combined to obtain the M groups of multi-frequency band speech signal feature vectors; The first path includes a first separable convolution module, a second separable convolution module, and a third separable convolution module, while the second path includes a fourth separable convolution module, a first depthwise separable convolution module, and a second depthwise separable convolution module.

4. The voice detection method according to claim 3, characterized in that, The step of inputting the M groups of multi-frequency band speech signals into the first path of the one-dimensional deep separable convolutional neural network to extract the time features of the multi-frequency band speech signals, and obtaining the time feature vector of the M groups of multi-frequency band speech signals, includes: The multi-band speech signal is input into the first separable convolution module to obtain the first multi-band speech signal feature vector, and the average value and variance of the first multi-band speech signal feature vector are calculated. The first multi-band speech signal feature vector is input into the second separable convolution module to obtain the second multi-band speech signal feature vector. The second multi-band speech signal feature vector is input into the third separable convolution module to obtain the third multi-band speech signal feature vector. The average value and variance of the first multi-band speech signal feature vector are combined with the third multi-band speech signal feature vector to form the time feature vector of the multi-band speech signal.

5. The voice detection method according to claim 3, characterized in that, The step of inputting the M groups of multi-frequency band speech signals into the second path of the one-dimensional deep separable convolutional neural network to extract the time-frequency features of the multi-frequency band speech signals, and obtaining the time-frequency feature vectors of the M groups of multi-frequency band speech signals, includes: The multi-band speech signal is input into the first depthwise separable convolution module to obtain the fourth multi-band speech signal feature vector, and the mean and variance of the fourth multi-band speech signal feature vector are calculated. The fourth multi-band speech signal feature vector is input into the fourth separable convolution module to obtain the fifth multi-band speech signal feature vector. The fifth multi-band speech signal feature vector is input into the second depthwise separable convolution module to obtain the sixth multi-band speech signal feature vector. The average value and variance of the fourth multi-band speech signal feature vector are combined with the sixth multi-band speech signal feature vector to form the time-frequency feature vector of the multi-band speech signal.

6. The voice detection method according to claim 1, characterized in that, The step of classifying the feature vector of the multi-band speech signal using a fully connected network to obtain a classification result, and using the classification result as the anomaly index of the speech signal to be detected, further includes: comparing the anomaly index of the speech signal to be detected with a preset threshold to determine the state of the speech signal to be detected.

7. A voice detection device, characterized in that, include: The filtering module is used to input the speech signal to be detected into an interpretable convolutional filter bank to obtain a multi-band speech signal; Specifically, it includes: The speech signal to be detected is divided into M frames; M frames of speech signals to be detected are respectively input into the interpretable convolutional filter group to obtain M groups of multi-frequency band speech signals; The feature extraction module is used to extract the time features of the multi-band speech signal using the first path of the one-dimensional deep separable convolutional neural network, extract the time-frequency features of the multi-band speech signal using the second path of the one-dimensional deep separable convolutional neural network, and combine the time features of the multi-band speech signal with the time-frequency features of the multi-band speech signal as a feature vector of the multi-band speech signal. The classification module is used to classify the feature vectors of the multi-band speech signal using a fully connected network to obtain a classification result, and uses the classification result as an anomaly index of the speech signal to be detected; specifically, it includes: The M groups of multi-band speech signal feature vectors are input into a fully connected network for classification to obtain M groups of classification results. The average of the M classification results is taken as the anomaly index of the speech signal to be detected.

8. The voice detection device according to claim 7, characterized in that, It also includes a comparison module, which compares the anomaly index of the voice signal to be detected with a preset threshold to determine the state of the voice signal to be detected.

Images