CN111239687A - Sound source positioning method and system based on deep neural network - Google Patents

Abstract

The invention discloses a positioning method, comprising: S1. acquiring a voice signal received by a microphone, and generating a voice data set; S2. preprocessing the voice signal in the voice data set; S3. calculating the phase of the sound source signal corresponding to the voice signal Weighted generalized cross-correlation function; S4. Obtain the time delay information corresponding to the peak of the phase-weighted generalized cross-correlation function, and use the time delay information as the TDOA observation value of the sound source signal reaching the microphone; Obtain the amplitude corresponding to the time delay information; S5. Use the TDOA The observation value and the amplitude are combined as the input vector, the three-dimensional space position coordinates corresponding to the sound source signal are used as the output vector, and the input vector and the output vector are combined to generate the feature vector; S6. Preprocess the feature vector; S7. Set the deep neural network correlation parameters of the training set, and train the deep neural network with the feature vector of the training set to obtain the trained deep neural network; S8. Input the input vector of the test set into the trained deep neural network for prediction, and obtain the three-dimensional spatial coordinates of the sound source signal.

Description

A sound source localization method and system based on deep neural network

technical field

The invention relates to the technical field of indoor sound source localization, in particular to a sound source localization method and system based on a deep neural network.

Background technique

In recent years, intelligent service products (such as smart speakers, smart homes, etc.) have been widely used in real life. In order to obtain a good user experience, the human-computer interaction capabilities of products have attracted more and more attention. In human-computer interaction, voice communication is an indispensable part. People can directly issue a voice password to command the machine to provide corresponding services. The machine recognizes the voice password and provides corresponding services without manual operation. At present, in near-field speech recognition application scenarios (such as mobile phones), the quality of the speech signal received by the microphone is very high, and the speech recognition rate has met the actual requirements. However, in far-field speech recognition application scenarios such as smart homes, the quality of the speech signal captured by the microphone is poor, and the speech recognition rate is low, which cannot meet the actual requirements. Therefore, solving the application of far-field speech recognition has increasingly become a research focus of domestic and foreign institutions in recent years. At present, the sound source localization algorithm is used in the front-end of speech recognition to estimate the position of the sound source, enhance the sound source signal in the direction of the position, and at the same time weaken the interference signals in other directions, which can improve the quality of the speech signal and improve the speech recognition rate. This method can effectively solve the problem. The application of far-field speech recognition has landed. Among them, effective sound source localization before speech recognition is of great significance in practice.

Classical localization algorithms are mainly sound source localization algorithms in two-dimensional space, which are mainly divided into three categories: one is the algorithm based on Time Difference of Arrival (TDOA). The time delay estimation algorithm, also known as the time difference of arrival algorithm, determines the position of the sound source according to the time difference between two microphones at different positions receiving the same sound source signal. It takes the time delay corresponding to the maximum peak of the Generalized Cross Correlation (GCC) of the sound signals received by the two microphones as the time delay estimation, and then uses the geometric constraints of the microphone array to obtain the sound source position estimation. This type of method is easily affected by environmental noise and indoor reverberation. When the noise is large or the reverberation is severe, there will be multiple spurious peaks in the GCC function, and it is easy to estimate the wrong TDOA value, which will lead to wrong sound sources. Location estimation. The second is an algorithm based on spatial spectrum estimation. The basic idea of the algorithm based on spatial spectrum estimation is to determine the direction angle and the position of the sound source according to the spatial spectrum. Since the estimation of the space signal is similar to the frequency estimation of the time domain signal, the estimation method of the space spectrum can be generalized by the nonlinear spectrum of the time domain, but the premise of this kind of algorithm is that the signal source is continuously distributed and the space is stationary, so this kind of algorithm The use is greatly restricted. One of the representative algorithms of the spatial spectrum algorithm is the feature subspace class algorithm. The feature subspace class algorithm can be divided into a subspace decomposition class algorithm and a subspace fitting class algorithm. The main algorithm of the former is Multiple Signal Classification (Multiple Signal Classification, MUSIC) algorithm and rotation invariant subspace algorithm (EstimatingSignal Parameter via Rotational Invariance Techniques, ESPRIT), the latter algorithm is mainly the maximum likelihood algorithm (Maximum Likelihood, ML) and weighted subspace fitting algorithm (Weighted SubspaceFitting, WSF) Wait. The third is the algorithm based on the controllable beam response. The algorithm based on the steerable beam response is to globally search the position of the maximum energy in the microphone array, which is the position of the sound source. Usually, the speech signal collected by the microphone is filtered, weighted and summed to form a beam, and then the point with the maximum output power of the beam is obtained, which is the position of the sound source. The algorithm based on the controllable beam response can be divided into the delay accumulation beam algorithm and the adaptive beam algorithm. Although the delay accumulation algorithm has small signal distortion and small calculation amount, its anti-interference ability is weak, and it is easily affected by noise. The adaptive algorithm has a large amount of calculation, and the signal has a certain distortion but has strong anti-interference ability.

Most of the sound source localization algorithms in three-dimensional space now use Multi-Modal fusion algorithms, of which the representative algorithm is the Audio-Visual Fusion algorithm. Usually, the face position information collected by the camera and the delay estimation (DOA) collected by the microphone are used to estimate the sound source position. This algorithm not only avoids the shortcomings of traditional image tracking that are limited by the number of cameras and light intensity, but also avoids The traditional sound source tracking is limited by the shortcomings of background noise and indoor reverberation, and the influence of environmental factors is greatly reduced. However, in the multi-modal fusion algorithm, we still need to set many parameters, and when the environment changes, the robustness of the algorithm decreases.

In recent years, the use of neural networks for sound source localization is a relatively popular research direction, especially after the development of deep learning. The research of sound source localization algorithm using neural network is usually to extract the feature vector from the speech signal first, and then transfer the feature vector to the neural network for training. The common speech feature vector is composed of TDOA of multiple microphone pairs, and does not use the amplitude information corresponding to the TDOA, and the amplitude corresponding to the TDOA reflects the reliability of the TDOA to a certain extent.

In general, the sound source localization method based on deep neural network is a research hotspot of indoor sound source localization, and this research is of great significance for solving many current audio applications, such as intelligent voice interaction technology. However, the current research on sound source localization methods based on deep neural networks is still immature, and the existing results have more or less certain deficiencies.

SUMMARY OF THE INVENTION

及其所对应的幅值

共同作为深度神经网络的输入向量，三维空间坐标作为深度神经网络的输出向量，适用于室内声源定位，具有良好的可扩展性以及算法鲁棒性。The purpose of the present invention is to provide a sound source localization method and system based on a deep neural network in view of the defects of the prior art.

and its corresponding amplitude

Together as the input vector of the deep neural network, and the three-dimensional space coordinates as the output vector of the deep neural network, it is suitable for indoor sound source localization, and has good scalability and algorithm robustness.

In order to achieve the above purpose, the present invention adopts the following technical solutions:

A sound source localization method based on a deep neural network, including a training phase of the deep neural network and a testing phase of the deep neural network, including steps:

S1. Acquire the voice signal received by the microphone, and generate a voice data set from the obtained voice signal; wherein, the voice data set includes a training data set and a test data set;

S2. carry out first preprocessing to the voice signal in the generated voice data set;

S3. calculate the phase-weighted generalized cross-correlation function of the sound source signal corresponding to the preprocessed speech signal;

S4. Obtain time delay information corresponding to the peak of the phase-weighted generalized cross-correlation function, and use the obtained time delay information as the TDOA observation value of the sound source signal arriving at the microphone; and obtain the amplitude corresponding to the time delay information value;

S5. the TDOA observation value and the amplitude are combined as the input vector of the deep neural network, the three-dimensional space position coordinates corresponding to the sound source signal are used as the output vector of the neural network, and the input vector and the output vector are combined to generate a feature vector;

S6. The second preprocessing is performed on the generated feature vector;

S7. In the training phase of the deep neural network, set the parameters related to the deep neural network, and train the deep neural network with the feature vector of the training set to obtain a trained deep neural network;

S8. In the test phase of the deep neural network, the input feature vector of the test set is transmitted to the trained deep neural network for prediction, the three-dimensional spatial position coordinates of the sound source signal are obtained, and the performance of the deep neural network model is evaluated by cross-validation.

Further, the set of microphone arrays in the step S1 is V={1,2,...,M}; each microphone node m includes two microphones, where m∈V; M represents a total of M pairs of microphones.

Further, the step S2 is specifically to perform first preprocessing on the speech signals received by the two microphones in the microphone node m, where the first preprocessing includes framing, windowing, and pre-emphasis.

Further, the step S3 is specifically calculating the phase-weighted generalized cross-correlation function R _m (τ) of the two microphone voice signals in the preprocessed microphone node m, which is expressed as:

和

分别表示为在节点m处的时域麦克风信号

和

的所对应的频域部分；符号*表示为复共轭操作。where m∈V,

and

are denoted as the time-domain microphone signal at node m, respectively

and

The corresponding frequency domain part of ; the symbol * denotes a complex conjugate operation.

表示为：Further, the step S4 obtains the time delay information corresponding to the peak of the phase weighted generalized cross-correlation function R _m (τ)

Expressed as:

对应的幅值

and obtain the delay information

corresponding amplitude

Further, the step S5 is specifically:

及其所对应的幅值

联合作为深度神经网络的输入向量I：Calculate the delay information of all nodes

and its corresponding amplitude

The union is the input vector I of the deep neural network:

The three-dimensional space position coordinate Q corresponding to the sound source signal S is used as the output vector of the neural network:

The input vector I and the output vector Q are combined to generate the feature vector G:

G=(I,Q) ^T .

Further, the second preprocessing in step S6 includes data cleaning, data disorder, and data normalization.

Further, the cross-validation used in the step S8 includes a leave-one-out validation method.

Correspondingly, a sound source localization system based on a deep neural network is also provided, including:

The first acquisition module is used to acquire the voice signal received by the microphone, and generate a voice data set from the acquired voice signal; wherein, the voice data set includes a training data set and a test data set;

a first preprocessing module, configured to perform a first preprocessing on the voice signal in the generated voice data set;

a calculation module for calculating the phase-weighted generalized cross-correlation function of the sound source signal corresponding to the preprocessed speech signal;

The second acquisition module is configured to acquire the time delay information corresponding to the peak of the phase-weighted generalized cross-correlation function, and use the acquired time delay information as the TDOA observation value of the sound source signal arriving at the microphone; and acquire the time delay information. The amplitude corresponding to the delay information;

The generation module is used to combine the TDOA observation value and the amplitude value as the input vector of the deep neural network, use the three-dimensional space position coordinates corresponding to the sound source signal as the output vector of the neural network, and combine the input vector and the output vector to generate Feature vector;

A second preprocessing module, configured to perform a second preprocessing on the generated feature vector;

The training module is used to set the parameters related to the deep neural network, and train the deep neural network with the feature vector of the training set to obtain the trained deep neural network

Further, it also includes:

The test module is used to input the input vector of the test set into the trained deep neural network for prediction, obtain the three-dimensional spatial position coordinates of the sound source signal, and use cross-validation to evaluate the performance of the deep neural network model.

及其所对应的幅值

共同作为深度神经网络的输入向量，三维空间坐标作为深度神经网络的输出向量，适用于室内声源定位，具有良好的可扩展性以及算法鲁棒性。Compared with the prior art, the present invention estimates the time delay

and its corresponding amplitude

Together as the input vector of the deep neural network, and the three-dimensional space coordinates as the output vector of the deep neural network, it is suitable for indoor sound source localization, and has good scalability and algorithm robustness.

Description of drawings

1 is a flowchart of a method for sound source localization based on a deep neural network provided by Embodiment 1;

2 is a schematic top view of a simulation environment provided by Embodiment 1, wherein a circle represents the position of a microphone;

3 is a flowchart of a training phase of the deep neural network provided by Embodiment 1;

FIG. 4 is a flow chart of the testing phase of the deep neural network provided in the first embodiment.

Detailed ways

The embodiments of the present invention are described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments may be combined with each other under the condition of no conflict.

The purpose of the present invention is to provide a sound source localization method and system based on a deep neural network in view of the defects of the prior art.

Example 1

The first embodiment provides a sound source localization method based on a deep neural network, including a training phase of the deep neural network and a testing phase of the deep neural network, as shown in Figure 1-2, including steps:

S11. Obtain the voice signal received by the microphone, and generate a voice data set from the obtained voice signal; wherein, the voice data set includes a training data set and a test data set;

S12. Perform first preprocessing on the voice signal in the generated voice data set;

S13. Calculate the phase-weighted generalized cross-correlation function of the sound source signal corresponding to the preprocessed speech signal;

S14. Acquire the time delay information corresponding to the peak of the phase-weighted generalized cross-correlation function, and use the obtained time delay information as the TDOA observation value of the sound source signal arriving at the microphone; and obtain the amplitude corresponding to the time delay information value;

S15. The TDOA observation value and the amplitude are combined as the input vector of the deep neural network, the three-dimensional space position coordinates corresponding to the sound source signal are used as the output vector of the neural network, and the input vector and the output vector are combined to generate a feature vector;

S16. Second preprocessing is performed on the generated feature vector;

S17. In the training phase of the deep neural network, set the parameters related to the deep neural network, and train the deep neural network with the feature vector of the training set to obtain a trained deep neural network;

S18. In the test phase of the deep neural network, the input vector of the test set is transmitted to the trained deep neural network for prediction, the three-dimensional spatial position coordinates of the sound source signal are obtained, and the performance of the deep neural network model is evaluated by cross-validation.

In this embodiment, the distributed microphone array is specifically described:

The specific simulation settings are as follows: the simulation environment is a typical conference room with a size of 4.1m×3.1m×3m, in which there are L=12 randomly distributed microphones in total. The distance between two microphones in each microphone node is Dm=0.6m. For simplicity, the location of the microphone is on a plane with a height of 1.75m. The speed of sound propagation is c=343 m/s. In this embodiment, the original voice signal without reverberation is a single-channel pure male English pronunciation with a sampling frequency of 16 kHz, and the frame length of the voice signal is 120 ms. The indoor reverberation time is T60=0.1s, the signal-to-noise ratio is SNR=20dB, and the number of Monte Carlo experiments is 50. The distributed microphone array has M microphone nodes in total, that is, a set of microphone nodes V={1,2,...,M}. Each microphone node m contains two microphones, where m ∈ V.

In step S11, a voice signal received by the microphone is acquired, and a voice data set is generated from the acquired voice signal; wherein, the voice data set includes a training data set and a test data set;

In this embodiment, the sound source position set is set in a plane with a height of 1.5m-1.7m, and H=24000 position sample sets are uniformly collected as the data set of the neural network. In the MATLAB simulation environment, the Image model is used to simulate the room impulse response, and then the original unreverberated speech signal is convolved with the room impulse response and Gaussian white noise is added to finally simulate the signal received by the microphone.

In step S12, a first preprocessing is performed on the speech signal in the generated speech data set.

Specifically, the first preprocessing is performed on the speech signals received by the two microphones in the microphone node m, where the first preprocessing includes framing, windowing, and pre-emphasis.

A rectangular window is used to add a window to the speech signal, and the formula of the window function ω(n) of the rectangular window is:

where N represents the length of the window function.

The formula for pre-emphasis is:

H(z)=1-αz ^-1

Among them, α represents the pre-emphasis coefficient, and the range is 0.9<α<1.0. In this embodiment, the length of the window function is the frame length, and the pre-emphasis coefficient α=0.97.

In step S13, a phase-weighted generalized cross-correlation function of the sound source signal corresponding to the preprocessed speech signal is calculated.

Specifically, the phase-weighted generalized cross-correlation function R _m (τ) of the two microphone speech signals in the preprocessed microphone node m is calculated, which is expressed as:

和

分别表示为在节点m处的时域麦克风信号

和

的所对应的频域部分；符号*表示为复共轭操作。在本实施例中，M＝6。where m∈V,

and

are denoted as the time-domain microphone signal at node m, respectively

and

The corresponding frequency domain part of ; the symbol * denotes a complex conjugate operation. In this embodiment, M=6.

In step S14, the time delay information corresponding to the peak of the phase-weighted generalized cross-correlation function is obtained, and the obtained time delay information is used as the TDOA observation value of the sound source signal arriving at the microphone; and the time delay information is obtained corresponding amplitude.

并将时延信息

作为声源信号S到达麦克风节点m的TDOA观测值，表示为：Obtain time delay information corresponding to the peak of the phase-weighted generalized cross-correlation function R _m (τ)

and the delay information

As the TDOA observation of the sound source signal S arriving at the microphone node m, it is expressed as:

和

表示节点m处包含的麦克风对与声源信息S的距离，c表示声音传播速度；||·||表示欧几里得范数。然后获取所述时延信息

(即TDOA观测值)对应的幅值

Among them, τ∈[-τ _max ,τ _max ], τ _max represents the theoretical maximum time delay (TDOA) of the sound source information S reaching the microphone node m, namely

and

Represents the distance between the microphone pair contained at the node m and the sound source information S, c represents the sound propagation speed; ||·|| represents the Euclidean norm. Then obtain the delay information

(i.e. TDOA observations) corresponding to the amplitude

TDOA positioning is a method of positioning using time difference. By measuring the time it takes for the signal to arrive at the monitoring station, the distance to the source of the signal can be determined. Using the distance from the signal source to each monitoring station (with the monitoring station as the center, the distance is the radius as a circle), the position of the signal can be determined. However, the absolute time is generally difficult to measure. By comparing the absolute time difference between the signal arriving at each monitoring station, a hyperbola with the monitoring station as the focus and the distance difference as the long axis can be drawn. The intersection of the hyperbola is the position of the signal.

In step S15, the TDOA observation value and the amplitude are combined as the input vector of the deep neural network, the three-dimensional space position coordinates corresponding to the sound source signal are used as the output vector of the neural network, and the input vector and the output vector are combined to generate Feature vector.

(即TDOA观测值)及其所对应的幅值

结合作为深度神经网络的输入向量I：Specifically: convert the delay information

(i.e. TDOA observations) and their corresponding amplitudes

Combined with the input vector I as a deep neural network:

The three-dimensional space position coordinate Q corresponding to the sound source signal S is used as the output vector of the neural network:

Combine the input vector I and the output vector Q to generate the feature vector G:

G=(I,Q) ^T .

In step S16, a second preprocessing is performed on the generated feature vector. The second preprocessing includes data cleaning, data disorder, and data normalization.

The normalization adopts the method of min-max normalization, and its conversion function is:

表示样本数据归一化后的结果。在经过神经网络的训练过后，应经过反归一化得出数据的值，在本实施例中为声源点的三维空间位置。Among them, g _min and g _max represent the minimum and maximum values in the sample feature vector G, respectively;

Indicates the result after normalizing the sample data. After the training of the neural network, the data value should be obtained through inverse normalization, which is the three-dimensional spatial position of the sound source point in this embodiment.

Among them, the denormalized transformation function is:

表示样本数据归一化后的结果，g为反归一化后的结果。Among them, g _min and g _max represent the minimum and maximum values in the sample feature vector G, respectively,

Indicates the result after normalization of sample data, and g is the result after de-normalization.

In step S17, in the training stage of the deep neural network, parameters related to the deep neural network are set, and the deep neural network is trained with the feature vector of the training set, so as to obtain a trained deep neural network.

In this embodiment, the number of neurons in the input layer of the deep neural network (DNN) is set to 12, and the number of neurons in the output layer is set to 3. The hidden layer is set to three layers, the number of neurons in the first hidden layer is 12, the activation function is tanh function, the number of neurons in the second hidden layer is 15, the activation function is tanh function, and the number of neurons in the third hidden layer is 3. The activation function is the tanh function.

In this embodiment, the loss function of the neural network is set as the mean squared error (MSE) between the real space position vector Q and the neural network prediction estimation vector P, which is expressed as:

where U is the total number of current neural network iteration datasets.

In step S18, in the test phase of the deep neural network, the input vector of the test set is transmitted to the trained deep neural network for prediction, the three-dimensional spatial position coordinates of the sound source signal are obtained, and cross-validation is used to evaluate the performance of the deep neural network model. performance.

The input vector of the test set is passed into the trained deep neural network, and the three-dimensional spatial position coordinates P=[px,py,pz] ^T of the sound source signal can be predicted, and the performance of the deep neural network model is evaluated by cross-validation.

In this embodiment, the total number of data sets is 24,000, and the performance of the neural network is tested by the cross-validation method. The cross-validation method adopts the leave-one-out validation method, that is, 4,000 sample points are left as the test set and 20,000 samples are left as the test set. As a training set, the tested data will be used as part of the training set in the next process, and this process will be repeated until no new sample data needs to be predicted.

A sound source localization method based on a deep neural network in this embodiment includes a training phase of the deep neural network and a testing phase of the deep neural network.

As shown in Figure 3, in the training phase of the deep neural network, steps S11-S17 are included.

As shown in FIG. 4 , in the testing phase of the deep neural network, steps S11-S16 and S18 are included.

It should be noted that, in the testing phase of this embodiment, a trained deep neural network is obtained based on the training phase, and then testing and positioning is performed.

以及R_m(τ)最大峰值所对应的幅值

共同作为深度神经网络的输入向量，三维空间坐标作为深度神经网络的输出向量，适用于室内声源定位，具有良好的可扩展性以及算法鲁棒性。Compared with the prior art, this embodiment uses the time delay estimation

and the amplitude corresponding to the maximum peak value of R _m (τ)

Together as the input vector of the deep neural network, and the three-dimensional space coordinates as the output vector of the deep neural network, it is suitable for indoor sound source localization, and has good scalability and algorithm robustness.

Embodiment 2

This embodiment provides a sound source localization system based on a deep neural network, including:

The first acquisition module is used to acquire the voice signal received by the microphone, and generate a voice data set from the acquired voice signal; wherein, the voice data set includes a training data set and a test data set;

a first preprocessing module, configured to perform a first preprocessing on the voice signal in the generated voice data set;

a calculation module for calculating the phase-weighted generalized cross-correlation function of the sound source signal corresponding to the preprocessed speech signal;

The second acquisition module is configured to acquire the time delay information corresponding to the peak of the phase-weighted generalized cross-correlation function, and use the acquired time delay information as the TDOA observation value of the sound source signal arriving at the microphone; and acquire the time delay information. The amplitude corresponding to the delay information;

The generation module is used to combine the TDOA observation value and the amplitude value as the input vector of the deep neural network, use the three-dimensional space position coordinates corresponding to the sound source signal as the output vector of the neural network, and combine the input vector and the output vector to generate Feature vector;

A second preprocessing module, configured to perform a second preprocessing on the generated feature vector;

The training module is used to set the parameters related to the deep neural network, and train the deep neural network with the feature vector of the training set to obtain the trained deep neural network

Further, it also includes:

The test module is used to input the input vector of the test set into the trained deep neural network for prediction, obtain the three-dimensional spatial position coordinates of the sound source signal, and use cross-validation to evaluate the performance of the deep neural network model.

It should be noted that a sound source localization system based on a deep neural network in this embodiment is similar to that in the first embodiment, and details are not described here.

以及R_m(τ)最大峰值所对应的幅值

共同作为深度神经网络的输入向量，三维空间坐标作为深度神经网络的输出向量，适用于室内声源定位，具有良好的可扩展性以及算法鲁棒性。Compared with the prior art, this embodiment uses the time delay estimation

and the amplitude corresponding to the maximum peak value of R _m (τ)

Together as the input vector of the deep neural network, and the three-dimensional space coordinates as the output vector of the deep neural network, it is suitable for indoor sound source localization, and has good scalability and algorithm robustness.

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention pertains can make various modifications or additions to the described specific embodiments or substitute in similar manners, but will not deviate from the spirit of the present invention or go beyond the definitions of the appended claims range.

Claims (10)

1. a sound source localization method based on deep neural network, is characterized in that, comprises the training phase of deep neural network and the testing phase of deep neural network, comprises steps: S1. Acquire the voice signal received by the microphone, and generate a voice data set from the obtained voice signal; wherein, the voice data set includes a training data set and a test data set; S2. carry out first preprocessing to the voice signal in the generated voice data set; S3. calculate the phase-weighted generalized cross-correlation function of the sound source signal corresponding to the preprocessed speech signal; S4. Obtain time delay information corresponding to the peak of the phase-weighted generalized cross-correlation function, and use the obtained time delay information as the TDOA observation value of the sound source signal arriving at the microphone; and obtain the amplitude corresponding to the time delay information value; S5. the TDOA observation value and the amplitude are combined as the input vector of the deep neural network, the three-dimensional space position coordinates corresponding to the sound source signal are used as the output vector of the neural network, and the input vector and the output vector are combined to generate a feature vector; S6. The second preprocessing is performed on the generated feature vector; S7. In the training phase of the deep neural network, set the parameters related to the deep neural network, and train the deep neural network with the feature vector of the training set to obtain a trained deep neural network; S8. In the test phase of the deep neural network, the input vector of the test set is transmitted to the trained deep neural network for prediction, the three-dimensional spatial position coordinates of the sound source signal are obtained, and the performance of the deep neural network model is evaluated by cross-validation. 2. A sound source localization method based on a deep neural network according to claim 1, wherein the set of microphone arrays in the step S1 is V={1,2,...,M}; each microphone Node m contains two microphones, where m ∈ V; M represents a total of M microphone nodes. 3. a kind of sound source localization method based on deep neural network according to claim 2, is characterized in that, described step S2 specifically is to carry out the first preprocessing to the speech signals received by two microphones in the microphone node m, The first preprocessing includes framing, windowing, and pre-emphasis. 4. a kind of sound source localization method based on deep neural network according to claim 2, is characterized in that, described step S3 is to calculate the phase-weighted generalized generalization of two microphone speech signals in the microphone node m after preprocessing specifically The cross-correlation function R _m (τ), expressed as:

where m∈V,

and

are denoted as the time-domain microphone signal at node m, respectively

and

The corresponding frequency domain part of ; the symbol * denotes a complex conjugate operation. 5. a kind of sound source localization method based on deep neural network according to claim 4, is characterized in that, described step S4 obtains the time delay corresponding to the peak of described phase-weighted generalized cross-correlation function R _m (τ) information

Expressed as:

and obtain the delay information

corresponding amplitude

6. a kind of sound source localization method based on deep neural network according to claim 5, is characterized in that, described step S5 is specifically: delay information

and its corresponding amplitude

Combined with the input vector I as a deep neural network:

The three-dimensional space position coordinate Q corresponding to the sound source signal S is used as the output vector of the neural network:

The input vector I and the output vector Q are combined to generate the feature vector G: G=(I,Q) ^T . 7 . The sound source localization method based on a deep neural network according to claim 6 , wherein the second preprocessing in the step S6 includes data cleaning, data disorder, and data normalization. 8 . 8 . The sound source localization method based on a deep neural network according to claim 7 , wherein the cross-validation adopted in the step S8 includes a leave-one-out validation method. 9 . 9. A sound source localization system based on a deep neural network, characterized in that, comprising: The first acquisition module is used to acquire the voice signal received by the microphone, and generate a voice data set from the acquired voice signal; wherein, the voice data set includes a training data set and a test data set; a first preprocessing module, configured to perform a first preprocessing on the voice signal in the generated voice data set; a calculation module for calculating the phase-weighted generalized cross-correlation function of the sound source signal corresponding to the preprocessed speech signal; The second acquisition module is configured to acquire the time delay information corresponding to the peak of the phase-weighted generalized cross-correlation function, and use the acquired time delay information as the TDOA observation value of the sound source signal arriving at the microphone; and acquire the time delay information. The amplitude corresponding to the delay information; The generation module is used to combine the TDOA observation value and the amplitude value as the input vector of the deep neural network, use the three-dimensional space position coordinates corresponding to the sound source signal as the output vector of the neural network, and combine the input vector and the output vector to generate Feature vector; A second preprocessing module, configured to perform a second preprocessing on the generated feature vector; The training module is used to set parameters related to the deep neural network, and train the deep neural network with the feature vector of the training set to obtain a trained deep neural network. 10. A deep neural network-based sound source localization system according to claim 9, characterized in that, further comprising: The test module is used to input the input vector of the test set into the trained deep neural network for prediction, obtain the three-dimensional spatial position coordinates of the sound source signal, and use cross-validation to evaluate the performance of the deep neural network model.

Images