A speech enhancement method and system that fuses vocoder and phase estimation

Abstract

The application discloses a speech enhancement method and system fusing a vocoder and phase estimation, and at least comprises the following steps: step S1: constructing and training a vocoder enhancement unit for processing input noisy speech and outputting enhanced speech information; step S2: constructing and training a phase enhancement unit for processing input noisy speech and outputting enhanced phase information; and step S3: constructing a post-processing module for receiving the enhanced phase information obtained in step S2 and the enhanced speech information obtained in step S1 and performing fusion processing to obtain a noise-reduced speech signal. The speech enhancement method and system fuses the vocoder and the phase estimation, improves the quality of a speech enhancement model generation model, solves a key problem of phase processing by using a phase decoder and an anti-winding phase loss function, and effectively improves the performance of the speech enhancement technology in a low signal-to-noise ratio scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech enhancement, and in particular to a speech enhancement method and system fusing a vocoder and phase estimation. BACKGROUND

[0002] In daily life, speech is one of the information that people contact most and use most frequently. Because of environmental noise or various factors in the network transmission process, speech signals inevitably have noise, such as the horn sound of a car when on the road or the current sound caused by network signal problems during transmission. With the rapid development of intelligent networks and information transmission, the continuous penetration of AI technology in the process of people's life convenience, the accuracy of speech information transmission is becoming more and more important to human daily life, and the optimization and improvement of speech enhancement model are becoming more and more important.

[0003] Speech enhancement technology aims to improve speech quality through various algorithms; by reducing noise to improve the speech damaged by noise, i.e. noise reduction, is the most important field in speech enhancement. According to the number of channels of speech signals, it can be divided into two research directions of single-channel and multi-channel, among which single-channel speech enhancement as the basis of other researches is an important link.

[0004] The traditional speech enhancement architecture is mainly based on the application of Fourier transform. The speech waveform in the time domain is generated into the amplitude spectrum and phase spectrum in the time-frequency domain after Fourier transform, and then processed and restored into the enhanced speech waveform through inverse Fourier transform. With the continuous improvement of digital signal processing theory, it has become an effective aid for traditional speech enhancement models, and has derived many classic algorithms, such as subspace speech enhancement method, Wiener filter method, spectral subtraction method, etc. These algorithms mostly use Fourier transform to process the frequency spectrum of noise and non-noise after transformation. For example, the traditional spectral subtraction method can separate the noise frequency spectrum component; or like Wiener filter method applies a filter in the time domain to achieve the purpose of filtering noise.

[0005] With the great improvement of computer computing power in the early 21st century, deep learning has become one of the most important application trends in the field of information processing. Speech enhancement has gradually been widely included as a supervised learning problem in the field of machine learning, and speech enhancement combined with neural networks has become the mainstream method in the current research field. Speech enhancement methods based on deep learning neural networks can be roughly divided into two categories: the first method is to convert the time-frequency signal into a time-frequency spectrum through short-time Fourier transform, and apply a neural network structure to map it to the spectrum of clean speech; the other method is to directly input the time-frequency signal, and use the learning ability of neural networks to map it to clean speech. In the early stage of deep learning research, speech enhancement was still limited to enhancing the amplitude spectrum only, without changing the noise phase spectrum, but such methods have certain limitations and pose some challenges and risks. Because there is a random structure in the phase component, and the phase is a major factor affecting the quality of speech, some traditional models that rely only on amplitude components for modeling ignore the phase component, which will face the crisis of possible speech distortion problems caused by enhancing the amplitude spectrum only.

[0006] As can be seen from the above, the prior art scheme is limited to enhancing the amplitude spectrum only, without changing the noise phase spectrum, and typical methods such as ideal ratio mask (IBM) and phase-sensitive mask (PSM) are used. The prior art scheme method at least has the following defects:

[0007] On the one hand, the traditional processing method needs to make a distribution assumption on the audio signal first. Once the assumption deviates greatly from the actual situation, the enhanced speech will be very different from the pure speech, which may cause distortion phenomenon to occur, making it difficult to meet the current communication requirements, thus giving birth to a series of speech enhancement models that fully play the advantages of various deep learning networks. However, due to the complex relationship between the various attributes of speech, it is difficult to balance quality and speed in the design process of the neural network model, i.e., the quality of speech generation and the speed of model training and speech processing.

[0008] On the other hand, as a signal in the time domain, the frequency spectrum refers to the representation of a time-domain signal in the frequency domain, which can be obtained by performing Fourier transform on the signal. The resulting results are two graphs with amplitude and phase as the vertical axis and frequency as the horizontal axis, namely amplitude spectrum and phase spectrum. Among them, the phase spectrum has continuity and correlation, and the phase information is affected by the spectral components. The difference in phase at different frequencies may have different effects on the speech signal. The traditional single-channel speech enhancement model usually enhances the amplitude spectrum of the noisy speech, but lacks processing of the phase spectrum, which affects the quality of the generated speech. In recent years, most studies have tried to refine the complex spectrum after enhancing the amplitude spectrum to enhance the phase spectrum, but the compensation effect between phase and amplitude greatly reduces the accuracy of phase prediction. Especially in a low signal-to-noise ratio environment, noise may mask the subtle changes in the speech signal, making it difficult to accurately extract phase information. Therefore, how to accurately extract phase information, significantly enhance the phase spectrum, and improve the performance of speech enhancement technology in low signal-to-noise ratio scenarios has become a severe challenge for the development of technology in this field.

[0009] Therefore, in view of the defects of the prior art, it is necessary to propose a technical solution to solve the technical problems existing in the prior art. SUMMARY

[0010] Therefore, it is necessary to provide a speech enhancement method and system that fuses vocoder and phase estimation, uses Mel-spectrum as acoustic features, and uses high-fidelity generative adversarial network HiFi-GAN as a vocoder to achieve a balance between speed and quality, thereby solving the balance between processing speed and generation quality caused by different model structures. At the same time, a single-channel speech enhancement model based on time-frequency domain is used to directly process phase information through a phase decoder combined with an anti-winding loss function, and the phase information is fused with the vocoder output to obtain the denoised speech signal, thereby solving the phase problem.

[0011] In order to solve the technical problems existing in the prior art, the technical solution of the present application is as follows:

[0012] A speech enhancement method fusing vocoder and phase estimation, at least comprising the following steps:

[0013] Step S1: constructing and training a vocoder enhancement unit for processing the input noisy speech and outputting enhanced speech information;

[0014] Step S2: constructing and training a phase enhancement unit for processing the input noisy speech and outputting enhanced phase information;

[0015] Step S3: a post-processing module is constructed to receive the enhanced phase information obtained in step S2 and the enhanced speech information obtained in step S1 and perform fusion processing to obtain a noise-reduced speech signal.

[0016] As a further improvement, in step S2, the phase enhancement unit is provided with at least a phase noise reduction module, which is implemented using a neural network.

[0017] As a further improvement, step S2 further includes the following steps:

[0018] Step S21: training the phase noise reduction module;

[0019] Step S22: using the trained phase noise reduction module to process the noisy speech to obtain enhanced phase information;

[0020] The phase noise reduction module includes at least a phase encoder, a two-stage convolutional enhancement transformer, and a phase decoder.

[0021] As a further improvement, step S21 includes:

[0022] Step S211: obtaining a large number of paired training sets of noisy speech and clean speech, and performing data preprocessing thereon;

[0023] Step S212: extracting features of the speech signal from the preprocessed data set as input to the model in step S23;

[0024] Step S213: training the phase noise reduction module by a supervised learning method, using an anti-winding loss function in the training process, and obtaining the corresponding optimal model through multiple iterations.

[0025] As a further improvement, in step S213, the loss function used in the training process is as follows:

[0026]

[0027] wherein, represents the value of the anti-winding loss function, and represent the instantaneous phase loss, the group delay loss, and the instantaneous angular frequency loss, respectively, which are defined as follows:

[0028]

[0029]

[0030]

[0031] wherein, X p represents the phase spectrum of the clean speech, represents the clean speech phase spectrum predicted by the model during training, is an anti-warping function, Δ DF and Δ DT represent the differential operators along the frequency and time axes, respectively, represents the expectation, ||·||1 represents the L1 norm.

[0032] As a further improvement, step S22 further comprises:

[0033] Step S221: the noisy speech is subjected to short-time Fourier transform to obtain the corresponding noisy phase spectrum and noisy amplitude spectrum, the noisy amplitude spectrum is discarded, and only the noisy phase spectrum is retained;

[0034] Step S222: the noisy phase spectrum obtained in step S221 is taken as the input of the phase encoder, and the output is in the compressed time-frequency domain representation;

[0035] Step S223: the time-frequency domain representation obtained in step S222 is processed by a two-stage convolution enhancement transformer;

[0036] Step S224: the time-frequency domain representation after processing in step S223 is taken as the input of the phase decoder, and the output is in the clean enhanced phase spectrum

[0037] As a further improvement, the vocoder enhancement unit at least comprises a mel-spectrum extraction module, a mel-spectrum domain noise reduction module, and a neural network vocoder module, wherein the mel-spectrum domain noise reduction module and the neural network vocoder module are implemented by using a neural network.

[0038] As a further improvement, step S1 comprises the following steps:

[0039] Step S11: obtain features from the training set, obtain a training set composed of noisy speech and corresponding clean speech in the same way as step S211, extract noisy mel-spectrum from the noisy speech and corresponding clean mel-spectrum from the corresponding clean speech by using the mel-spectrum extraction module;

[0040] Step S12: use the mel-spectrum extracted in step S11 to train the mel-spectrum domain noise reduction module;

[0041] Step S13: use the clean speech obtained in step S11 and the extracted mel-spectrum to train the neural network vocoder module, and complete the training of the vocoder enhancement unit;

[0042] Step S14: use the trained vocoder enhancement unit to process the noisy speech to obtain the enhanced synthesized speech waveform.

[0043] The application further discloses a speech enhancement system fusing a vocoder and phase estimation, comprising at least a vocoder enhancement unit, a phase enhancement unit and a post-processing module, wherein,

[0044] The vocoder enhancement unit is used for processing the input noisy speech and outputting enhanced speech information.

[0045] The phase enhancement unit is used for processing the input noisy speech and outputting enhanced phase information.

[0046] The post-processing module is used for receiving the enhanced phase information obtained in step S2 and the enhanced speech information obtained in step S1 and performing fusion processing to obtain a speech signal after noise reduction.

[0047] As a further improved scheme, the phase enhancement unit is provided with at least a phase noise reduction module, and the phase noise reduction module comprises at least a phase encoder, a two-stage convolution enhancement transformer and a phase decoder.

[0048] The vocoder enhancement unit comprises at least a mel spectrum extraction module, a mel spectrum domain noise reduction module and a neural network vocoder module, wherein the mel spectrum domain noise reduction module and the neural network vocoder module are implemented by using a neural network; and the neural network vocoder module adopts a high-fidelity generative adversarial network (HiFi-GAN).

[0049] Compared with the prior art, the technical scheme of the application adopts a speech enhancement joint architecture fusing a vocoder and phase estimation, wherein the vocoder is a high-quality and fast vocoder HiFi-GAN, which balances the quality and speed to a certain extent, and multiple loss functions are used for optimization, further improving the quality of the speech enhancement model generation model; in addition, the phase decoder and the anti-winding phase loss function solve the key problem of phase processing, effectively improving the performance of the speech enhancement technology in a low signal-to-noise ratio scene. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 A flowchart of a speech enhancement method fusing a vocoder and phase estimation is provided for the application.

[0051] Figure 2 A whole framework diagram of a speech enhancement system fusing a vocoder and phase estimation is provided for the application.

[0052] Figure 3 A structure block diagram of a phase noise reduction module is provided for the application.

[0053] Figure 4 A schematic diagram of a generator structure in a neural network vocoder module is provided for the application.

[0054] Figure 5The schematic diagram of the discriminator structure in the neural network vocoder module of the application.

[0055] Figure 6 The structural schematic diagram of the mel-frequency spectrum domain noise reduction module in the application.

[0056] Figure 7 The flow schematic diagram of the post-processing module in the application.

[0057] The following specific embodiments will further illustrate the application in combination with the above-mentioned drawings. DETAILED DESCRIPTION

[0058] The technical solutions provided by the application will be further illustrated in combination with the drawings.

[0059] Referring to Figure 1 , a flow block diagram of a speech enhancement method fusing a vocoder and phase estimation is shown, and at least includes the following steps:

[0060] Step S1: constructing and training a vocoder enhancement unit, for processing the input noisy speech and outputting enhanced speech information;

[0061] Step S2: constructing and training a phase enhancement unit, for processing the input noisy speech and outputting enhanced phase information;

[0062] Step S3: constructing a post-processing module, for receiving the enhanced phase information obtained in step S2 and the enhanced speech information obtained in step S1 and performing fusion processing to obtain a noise-reduced speech signal.

[0063] Referring to Figure 2 , a whole framework diagram of a speech enhancement system fusing a vocoder and phase estimation is shown, and at least includes a vocoder enhancement unit, a phase enhancement unit and a post-processing module, wherein,

[0064] The vocoder enhancement unit is used for processing the input noisy speech and outputting enhanced speech information;

[0065] The phase enhancement unit is used for processing the input noisy speech and outputting enhanced phase information;

[0066] The post-processing module is used for receiving the enhanced phase information obtained in step S2 and the enhanced speech information obtained in step S1 and performing fusion processing to obtain a noise-reduced speech signal.

[0067] Based on the architecture, in step S2, the phase enhancement unit at least sets a phase noise reduction module, the phase noise reduction module is implemented by using a neural network and includes the following steps:

[0068] Step S21: training the phase noise reduction module;

[0069] Step S22: processing the noisy speech using the trained phase denoising module to obtain enhanced phase information;

[0070] Step S21 further comprises:

[0071] Step S211: obtaining a large number of paired training sets of noisy speech and clean speech, and pre-processing the training sets;

[0072] Step S212: extracting features of the speech signal from the pre-processed data set for training of the model in step S213;

[0073] Step S213: training the phase denoising module by a supervised learning method, using an anti-winding loss function in the training process, and obtaining the corresponding optimal model through multiple iterations; wherein the loss function is shown in formula 1 as follows:

[0074]

[0075] wherein, represents the value of the anti-winding loss function, and respectively represent the instantaneous phase loss, the group delay loss and the instantaneous angular frequency loss, which are respectively defined as:

[0076]

[0077]

[0078]

[0079] wherein, X p represents the phase spectrum of the clean speech (obtained from step 2122), represents the phase spectrum of the clean speech predicted by the model during the training process, is an anti-winding function used to avoid the error propagation problem caused by phase winding, and Δ DF and Δ DT respectively represent the differential operators along the frequency axis and the time axis, represents the expectation, and ||·||1 represents the L1 norm.

[0080] In the above technical solution, step S211 further comprises:

[0081] Step S2111: using a multi-language data set Voice Bank and a noise data set DEMAND to simulate a noisy speech data set, using a clean speech data set of Voice Bank, matching the noisy speech and the corresponding clean speech in pairs to form a training set;

[0082] Step S2112: downsample all the speech in the training set to 16 kHz;

[0083] Step S212 further comprises:

[0084] Step S2121: obtain noisy speech phase spectrum from noisy speech dataset by short-time Fourier transform;

[0085] Step S2122: obtain clean speech phase spectrum from clean speech dataset by short-time Fourier transform;

[0086] Step S2123: use the speech phase spectrum obtained from step S2121 and step S2122 for training of step S213 model;

[0087] As shown in Figure 3 , the phase denoising module at least comprises a phase encoder, a two-stage convolution enhanced transformer and a phase decoder, wherein step S22 further comprises:

[0088] Step S221: after the noisy speech is subjected to short-time Fourier transform, the corresponding noisy phase spectrum and noisy amplitude spectrum are obtained, the noisy amplitude spectrum is discarded, and only the noisy phase spectrum is retained;

[0089] Step S222: the noisy phase spectrum obtained in step S221 is taken as the input of the phase encoder, and the output is in the compressed time-frequency domain representation;

[0090] Step S223: the time-frequency domain representation obtained in step S222 is processed by a two-stage convolution enhanced transformer;

[0091] Step S224: the time-frequency domain representation after the processing in step S223 is taken as the input of the phase decoder, and the output is in the clean enhanced phase spectrum

[0092] Wherein, step S222 further comprises:

[0093] Step S2221: after the convolution block 1 is activated by the activation function PReLu (formula 5), the noisy phase spectrum is taken as the feature, and the feature dimension is increased by increasing the number of channels in the convolution layer;

[0094]

[0095] Wherein, y i is the input of the nonlinear activation function f of the i-th channel, a i is the coefficient corresponding to the control negative part slope, which is a learnable parameter.

[0096] Step S2222: The dilated dense network uses four convolutional layers with dilation sizes of 1, 2, 4, and 8 to expand the receptive field along the time axis, and dense connections are used for all convolutional layers to avoid the problem of gradient disappearance;

[0097] Step S2223: Finally, the convolutional block 2 down-samples the features by increasing the step size of the convolutional layer;

[0098] Step S224 further includes:

[0099] Step S2241: The dilated dense network connects the convolutional layers, and the deconvolutional block up-samples the features;

[0100] Step S2242: Then, the parallel phase estimation architecture (in the dashed box) uses two parallel 2D convolutional layers to output the real part and the imaginary part of the phase Figure 3 respectively. The clean phase spectrum is predicted by a two-argument arctangent function (Arctan2) (formula 6)

[0101]

[0102] where Sgn * (t) is equal to 1 when t≥0; and Sgn * (t) is equal to -1 when t<0.

[0103] Referring to Figure 2 , the vocoder enhancement unit at least includes a mel-spectrum extraction module, a mel-spectrum domain noise reduction module, and a neural network vocoder module, wherein the mel-spectrum domain noise reduction module and the neural network vocoder module are implemented by using a neural network; step S1 includes the following steps:

[0104] Step S11: Obtain features from a training set, obtain a training set composed of noisy speech and corresponding clean speech in the same manner as step S211, extract noisy mel-spectra from the noisy speech and corresponding clean mel-spectra from the corresponding clean speech by using the mel-spectrum extraction module.

[0105] Step S12: Train the mel-spectrum domain noise reduction module using the mel-spectra extracted in step S11;

[0106] Step S13: Train the neural network vocoder module using the clean speech obtained in step S11 and the extracted mel-spectra, and complete the training of the vocoder enhancement unit;

[0107] Step S14: Process the noisy speech using the trained vocoder enhancement unit to obtain enhanced synthesized speech waveforms;

[0108] ​The neural network vocoder module employs a generative adversarial network structure; see [link / reference]. Figure 4 The diagram shows the generator structure in the neural network vocoder module. More specifically, the neural network vocoder is a High-Fidelity Generative Adversarial Network (HiFi-GAN), and its constituent module is a generator (such as...). Figure 4 ) and two discriminators (see Figure 5 ,in, Figure 5 'a' represents the multi-period discriminant (MPD). Figure 5 b is the multi-scale discriminator (MSD). During training, the generator and the dual discriminator engage in an adversarial game to gradually optimize the model parameters. The training steps include:

[0109] Step S131: Fix the generator and train the discriminator: as follows Figure 5 a. This figure uses the second sub-discriminator of MPD with a cycle number of 3 as an example. The speech input to MPD is transformed from one-dimensional to two-dimensional, and arranged into 3 groups of data with the same cycle number, with the length shortened to one-third of the original. Then, each group of data with the same cycle is activated by the Leaky ReLU activation function (as in Equation 5, the difference being a). i A convolutional layer activated with a fixed parameter is used, with the kernel size set to 1 to process each cycle separately, ultimately determining the authenticity of the speech waveform; while... Figure 5 Figure b illustrates the second sub-discriminator of the MSD as an example. The input speech data to the MSD passes through an average pooling layer, then through strided and grouped convolutional layers activated by leaky ReLU, finally determining the authenticity of the speech waveform. The discriminator size is adjusted by modifying the number of convolutional layers and the stride.

[0110] Step S132: Fix the discriminator and train the generator: First, the Mel spectrum is sampled |ku| times in a loop through a deconvolution layer to obtain a dimension-matched sequence. An MRF (Multi-receptive field fusion) structure is bound to this deconvolution layer. This structure consists of multiple ResBlock modules, each containing a convolutional layer activated by Leaky ReLU. Each ResBlock has a different kernel size and dilation rate to process and extract input and data packets of different lengths. After multiple loops, the corresponding speech waveform is generated.

[0111] The three loss functions used in step S13 are: the generative adversarial network loss function (Equations 7 and 8), the Mel spectrum loss function (Equation 9), and the feature matching loss function (Equation 10), which are defined as follows:

[0112]

[0113]

[0114] wherein, represents the loss function of the discriminator when fixing the generator and training the discriminator, represents the loss function of the generator when fixing the discriminator and training the generator, x is the real audio input, and s is the mel-frequency input of the real audio, represents the expectation, D(·) represents the judgment result of the discriminator D on the input, and G(·) represents the generation result of the generator G on the input.

[0115]

[0116] wherein, φ is a mapping function of the speech waveform to the corresponding mel spectrum.

[0117]

[0118] wherein, T is the number of layers of the discriminator, D i and N i are the features and the number of features of the i-th layer of the discriminator, respectively.

[0119] The step S14 further comprises:

[0120] Step S141: The mel spectrum extraction module extracts the noisy mel spectrum from the noisy speech.

[0121] Step S142: The noisy mel spectrum extracted in step S141 is input to the mel spectrum domain noise reduction module, and after processing, the enhanced mel spectrum is output. Referring to Figure 6 , a structure diagram of the mel spectrum domain noise reduction module is shown.

[0122] Step S143: The enhanced mel spectrum obtained in step S142 is input to the neural network vocoder module, and after processing, the enhanced speech synthesis waveform is output.

[0123] In step S3, the post-processing module will use the enhanced phase spectrum output by the phase enhancement unit to process the synthesized speech waveform output by the vocoder enhancement unit, thereby outputting the noise-reduced speech signal. Referring to Figure 7 , a flowchart of the post-processing module is shown, and the steps thereof include:

[0124] Step S31: The synthesized speech waveform and the enhanced phase spectrum obtain the enhanced complex spectrum;

[0125] Step S32: The enhanced complex spectrum undergoes a series of processing;

[0126] Step S33: Finally, the signal is recovered to time domain by inverse short-time Fourier transform (ISTFT), and the denoised speech waveform is output.

[0127] The above description of the embodiments is only used to help understand the method of the present application and its core idea. It should be noted that, for those skilled in the art, some improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.

[0128] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A speech enhancement method integrating vocoder and phase estimation, characterized in that, It should include at least the following steps: Step S1: Construct and train the vocoder enhancement unit to process the input noisy speech and output enhanced speech information; Step S2: Construct and train a phase enhancement unit to process the input noisy speech and output enhanced phase information; Step S3: Construct a post-processing module to receive the enhanced phase information obtained in step S2 and the enhanced speech information obtained in step S1 and perform fusion processing to obtain the noise-reduced speech signal; The vocoder enhancement unit includes at least a Mel spectrum extraction module, a Mel spectrum domain noise reduction module, and a neural network vocoder module, wherein the Mel spectrum domain noise reduction module and the neural network vocoder module are implemented using neural networks; step S1 includes the following steps: Step S11: Obtain features from the training set to obtain a training set consisting of noisy speech and corresponding clean speech. Extract noisy Mel spectra from noisy speech and corresponding clean Mel spectra from corresponding clean speech using the Mel spectrum extraction module. Step S12: Use the Mel spectrum extracted in step S11 to train the Mel spectrum domain noise reduction module; Step S13: Using the clean speech and extracted Mel spectrum obtained in step S11, train the neural network vocoder module to complete the training of the vocoder enhancement unit. Step S14: Use the trained vocoder enhancement unit to process the noisy speech and obtain the enhanced synthesized speech waveform; In step S2, the phase enhancement unit shall at least include a phase denoising module, which shall be implemented using a neural network; step S2 also includes the following steps: Step S21: Train the phase noise reduction module; Step S22: Use the trained phase denoising module to process the noisy speech and obtain the enhanced phase information; The phase noise reduction module includes at least a phase encoder, a two-stage convolutional enhancement transformer, and a phase decoder; step S21 includes: Step S211: Obtain a large training set of noisy and clean speech pairs and preprocess the data. Step S212: Extract features of the speech signal from the preprocessed dataset and use them as input to the model in step S23; Step S213: Train the phase denoising module using supervised learning. During training, an anti-wrap loss function is used, and the optimal model is obtained through multiple iterations. The loss function used in step S213 during training is as follows: （1） in, This represents the value of the anti-winding loss function. and These represent instantaneous phase loss, group delay loss, and instantaneous angular frequency loss, respectively, and are defined as follows: （2） （3） （4） in, Represents the phase spectrum of clean speech. This represents the clean speech phase spectrum predicted by the model during training. , It is an anti-winding function. and These represent the differential operators along the frequency axis and the time axis, respectively. Expressing expectations, Representing the L1 norm; step S22 further includes: Step S221: After the noisy speech undergoes a short-time Fourier transform, the corresponding noisy phase spectrum and noisy amplitude spectrum are obtained. The noisy amplitude spectrum is discarded, and only the noisy phase spectrum is retained. Step S222: The noisy phase spectrum obtained in step S221 is used as the input of the phase encoder, and the output is represented by its compressed time-frequency domain. Step S223: Process the time-frequency domain representation obtained in step S222 using a two-stage convolutional enhancement transformer; Step S224: The time-frequency domain representation, after processing in step S223, is used as the input to the phase decoder, and the output is a clean, enhanced phase spectrum. .

2. A speech enhancement system that integrates a vocoder and phase estimation using the method described in claim 1, characterized in that, It includes at least a vocoder enhancement unit, a phase enhancement unit, and a post-processing module, wherein, The vocoder enhancement unit is used to process the input noisy speech and output enhanced speech information. The phase enhancement unit is used to process the input noisy speech and output enhanced phase information. The post-processing module is used to receive the enhanced phase information obtained in step S2 and the enhanced speech information obtained in step S1 and perform fusion processing to obtain the noise-reduced speech signal. The phase enhancement unit shall be equipped with at least a phase noise reduction module, which shall include at least a phase encoder, a two-stage convolutional enhancement transformer and a phase decoder; The vocoder enhancement unit includes at least a Mel spectrum extraction module, a Mel spectrum domain noise reduction module, and a neural network vocoder module. The Mel spectrum domain noise reduction module and the neural network vocoder module are implemented using neural networks. The neural network vocoder module uses a high-fidelity generative adversarial network (HiFi-GAN).

Images