A speech enhancement method based on harmonic model fundamental frequency synchronization and improved Wiener filtering

Abstract

The present invention belongs to the technical field of speech enhancement, and specifically discloses a speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering, which calculates the response function of the Wiener filter according to the minimum mean square error criterion, obtains the power spectrum of the predicted speech signal output by the Wiener filter after the noisy speech signal passes through the Wiener filter; obtains the fundamental frequency information of the predicted speech signal through the fundamental frequency estimation algorithm; after the predicted speech signal is subjected to windowing processing and fundamental frequency synchronization Fourier transform, the voiced sound frame is subjected to harmonic modeling to obtain the reconstructed phase information; finally, the amplitude spectrum obtained by square rooting the power spectrum of the predicted speech signal and the reconstructed phase information are synthesized to enhance the speech signal. The method of the present invention introduces a harmonic model to reconstruct the phase information, and no longer uses the phase information of the noisy speech, and based on fundamental frequency synchronization, introduces the fundamental frequency parameter into the Fourier transform, abandons the fixed Fourier transform length and the fixed window length, so as to achieve a better speech denoising and enhancement effect.

Description

Technical Field

[0001] The invention belongs to the technical field of speech enhancement, and in particular relates to a speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering. Background Art

[0002] Speech enhancement methods include spectral subtraction and Wiener filtering algorithms.

[0003] Spectral subtraction converts the noisy speech signal into the frequency domain to obtain the power spectrum of the noisy speech signal, then uses the noise estimation algorithm to obtain the power spectrum of the noise signal in the noisy speech, and subtracts the power spectrum of the noise signal from the power spectrum of the noisy speech signal to obtain the power spectrum of the enhanced speech signal, that is, the power spectrum of the clean speech signal estimation, and then takes the square root to obtain the amplitude spectrum of the enhanced speech signal, and finally directly extracts the phase information of the noisy speech signal, and restores the time domain signal through inverse Fourier transform to obtain the enhanced speech signal. However, spectral subtraction has the following two disadvantages:

[0004] First, directly subtracting the power spectrum of the noise signal from the power spectrum of the noisy speech signal will have a residual power spectrum component, which appears as a spike pulse in the frequency domain and a superposition of sinusoidal signals in the time domain. It exhibits musical characteristics and is called musical noise. Second, when calculating the power spectrum from the time domain to the frequency domain, the phase information will inevitably be lost. However, the spectral subtraction method directly extracts the phase of the noisy speech signal as the phase of the enhanced speech signal in the synthesis stage, without reconstructing the accurate phase information.

[0005] The Wiener filtering algorithm is based on spectral subtraction and removes the residual noise of spectral subtraction by adding a Wiener filter.

[0006] The Wiener filtering algorithm converts music noise into white noise by adding a Wiener filter. Although it effectively suppresses music noise, it is also unable to perform phase reconstruction, and the enhanced speech signal obtained will still have white noise residual. Summary of the invention

[0007] The purpose of the present invention is to propose a speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering. The method is improved based on the Wiener filtering algorithm. By introducing a harmonic model for phase reconstruction, the phase information of noisy speech is no longer used. The idea of ​​fundamental frequency synchronization is introduced, and the fundamental frequency parameters are introduced into the Fourier transform. The fixed Fourier transform length and the fixed window length are abandoned to achieve better speech denoising and enhancement effect.

[0008] In order to achieve the above object, the present invention adopts the following technical scheme:

[0009] A speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering, comprising the following steps:

[0010] Step 1. Calculate the response function of the Wiener filter according to the minimum mean square error criterion to obtain the power spectrum of the predicted speech signal output by the noisy speech signal after passing through the Wiener filter;

[0011] Step 2. Obtain the fundamental frequency information of the predicted speech signal through a fundamental frequency estimation algorithm;

[0012] Step 3. Perform windowing processing and frequency synchronization Fourier transform on the predicted speech signal to obtain the Fourier transform form of the predicted speech signal;

[0013] Step 4. Perform harmonic modeling on the voiced frames in the predicted speech signal, and obtain reconstructed phase information based on the harmonic model;

[0014] Step 5. Root the power spectrum of the predicted speech signal output by the Wiener filter in step 1 to obtain the amplitude spectrum of the enhanced speech signal, and then combine it with the reconstructed phase information in step 4 to synthesize the enhanced speech signal.

[0015] In addition, based on the speech enhancement method based on the harmonic model fundamental frequency synchronization improved Wiener filter, the present invention also proposes a speech enhancement system based on the harmonic model fundamental frequency synchronization improved Wiener filter adapted thereto, and its technical scheme is as follows:

[0016] A speech enhancement system based on harmonic model fundamental frequency synchronization improved Wiener filtering, comprising:

[0017] The Wiener filter module is used to calculate the response function of the Wiener filter according to the minimum mean square error criterion, and obtain the power spectrum of the predicted speech signal output by the noisy speech signal after passing through the Wiener filter;

[0018] A fundamental frequency estimation module, used for obtaining fundamental frequency information of a predicted speech signal through a fundamental frequency estimation algorithm;

[0019] A Fourier transform module is used to perform windowing processing and baseband synchronization Fourier transform on the predicted speech signal to obtain a Fourier transform form of the predicted speech signal;

[0020] A phase reconstruction module, used for performing harmonic modeling on the voiced frames in the predicted speech signal, and obtaining reconstructed phase information based on the harmonic model;

[0021] And an enhanced speech synthesis module is used to obtain the amplitude spectrum of the enhanced speech signal by square rooting the power spectrum of the predicted speech signal output by the Wiener filter, and then synthesize the enhanced speech signal in combination with the reconstructed phase information of the phase reconstruction module.

[0022] In addition, based on the above-mentioned speech enhancement method based on harmonic model fundamental frequency synchronization to improve Wiener filtering, the present invention also proposes a computer device, which includes a memory and one or more processors; the memory stores executable code, and when the processor executes the executable code, it is used to implement the steps of the above-mentioned speech enhancement method based on harmonic model fundamental frequency synchronization to improve Wiener filtering.

[0023] In addition, based on the above-mentioned speech enhancement method based on harmonic model fundamental frequency synchronization to improve Wiener filtering, the present invention also proposes a computer-readable storage medium on which a program is stored; when the program is executed by a processor, it is used to implement the steps of the above-mentioned speech enhancement method based on harmonic model fundamental frequency synchronization to improve Wiener filtering.

[0024] The present invention has the following advantages:

[0025] As described above, the present invention discloses a speech enhancement method based on improved Wiener filtering with fundamental frequency synchronization of a harmonic model. The method introduces a harmonic model to reconstruct phase information and no longer uses the phase information of noisy speech. When the noise mixed with the noisy speech is white noise, other speech enhancement methods only filter out the white noise. The speech enhancement method of the present invention introduces a harmonic model to perform harmonic modeling on the voiced frames in the speech signal. It can accurately distinguish between unvoiced frames and voiced frames in the speech signal, and filter out white noise and unvoiced frames without semantic information, so as to achieve better speech denoising and enhancement effect. In addition, the present invention also introduces the idea of ​​fundamental frequency synchronization, abandons the fixed Fourier transform length and the fixed window length, and introduces the fundamental frequency parameter into the Fourier transform. When the noise mixed with the noisy speech is the speech of someone else, it is displayed as a voiced frame with slightly weaker energy in the time domain. Other speech enhancement methods protect all voiced frames, and therefore cannot accurately distinguish whether the voiced frames belong to the speaker, which leads to the presence of the noise of someone else's speech in the filtering result. The present invention introduces fundamental frequency synchronization in the Fourier transform process, and determines the speaker information according to the fundamental frequency information, so that the speech of someone else in the speech signal, which is a voiced frame, can be accurately filtered out as noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of a speech enhancement method based on harmonic model fundamental frequency synchronization to improve Wiener filtering in an embodiment of the present invention;

[0027] Figure 2 is a time domain diagram of a clean speech signal in an embodiment of the present invention;

[0028] Figure 3 : is a time domain diagram of a noisy speech signal containing 10 dB white noise in an embodiment of the present invention;

[0029] Figure 4is a time domain diagram of an enhanced speech signal obtained by spectral subtraction in an embodiment of the present invention;

[0030] Figure 5 is a time domain diagram of an enhanced speech signal obtained by the Wiener filtering algorithm in an embodiment of the present invention;

[0031] Figure 6 It is a time domain diagram of an enhanced speech signal obtained by the speech enhancement method of the present invention based on the harmonic model fundamental frequency synchronization improved Wiener filtering;

[0032] Figure 7 is a spectrum diagram of a clean speech signal in an embodiment of the present invention;

[0033] Figure 8 : is a spectrum diagram of a noisy speech signal containing 10 dB white noise in an embodiment of the present invention;

[0034] Fig. 9 is a spectrum diagram of an enhanced speech signal obtained by spectral subtraction in an embodiment of the present invention;

[0035] Fig.10 is a spectrum diagram of an enhanced speech signal obtained by the Wiener filtering algorithm in an embodiment of the present invention;

[0036] Fig.11 The spectrum diagram of the enhanced speech signal is obtained by the speech enhancement method of the present invention based on the harmonic model fundamental frequency synchronization improved Wiener filtering. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0038] Example 1

[0039] like Figure 1 As shown, the speech enhancement method based on harmonic model fundamental frequency synchronization to improve Wiener filtering in this embodiment includes the following steps:

[0040] Step 1. Calculate the response function of the Wiener filter according to the minimum mean square error criterion to obtain the noisy speech signal The predicted speech signal output by the Wiener filter The power spectrum of .

[0041] Assume that the noisy speech signal is , the clean speech signal is , the noise signal is , and Independence, three things are met:

[0042] ;

[0043] in is the time series of discrete speech signals.

[0044] Noisy speech signal The predicted speech signal after the Wiener filter is , then:

[0045] ;

[0046] in, Represents the time index, represents the Wiener filter order, represents the noisy speech signal at time nm, is the response function of the Wiener filter.

[0047] Error signal satisfy:

[0048] ;

[0049] Mean square error E [ ( ε ) 2 ] for:

[0050] E [ ( ε ) 2 ] = E [ ( s ( n ) − ∑ m ∈ p h ( m ) x ( n − m ) ) 2 ] ;

[0051] According to the minimum mean square error criterion, Taking partial derivatives and setting them equal to zero gives the covariance form:

[0052] ;

[0053] in, represents the frame index, Represents a clean speech signal With noisy speech signal The cross-covariance function of represents the response function of the Wiener filter, and The meaning is the same, but the context is different, so the meaning is different; Represents a noisy speech signal The autocovariance function of .

[0054] Then perform a short-time Fourier transform and get:

[0055] ;

[0056] in, and are the power spectrum of the clean speech signal and the power spectrum of the noisy speech signal, respectively. is the gain of the Wiener filter, represents the imaginary unit, Represents the angular frequency.

[0057] Predicting speech signals The power spectrum It is expressed as:

[0058] .

[0059] Step 2. Obtain the predicted speech signal through the fundamental frequency estimation algorithm Base frequency information.

[0060] Baseband It is the derivative of the fundamental frequency period. The fundamental frequency corresponds to the frequency of vocal cord vibration. The faster the vocal cords vibrate, the higher the fundamental frequency. It is an important characteristic of the speech excitation source. The typical fundamental frequency range of adult males is about 100-250 Hz, while the fundamental frequency range of adult females is about 150-350 Hz. Generally speaking, women have a slightly higher pitch than men.

[0061] The fundamental frequency estimation algorithms include the short-time autocorrelation method, the short-time average amplitude difference function AMDF method, etc.

[0062] In this embodiment, a pitch estimation filter PEFAC with amplitude compression is used. The pitch estimation filter can provide reliable pitch estimation and accurate unvoiced / voiced sound detection even under a low signal-to-noise ratio.

[0063] The fundamental frequency estimation algorithm is a prior art, and thus the present invention will not elaborate on it in detail.

[0064] Step 3. Predict the speech signal Perform windowing and frequency synchronization Fourier transform to obtain the Fourier transform form of the predicted speech signal .

[0065] Based on the short-time stationary characteristics of speech signals and the requirements of short-time Fourier transform, it is necessary to first predict the speech signal Perform frame processing, and then perform windowing processing on the subframes obtained by the frame processing.

[0066] The purpose of windowing is to eliminate the signal discontinuity at both ends of the subframe, so that both ends of the subframe gradually become zero, which is conducive to short-time Fourier transform and makes the spectrum more delicate.

[0067] During the windowing process, the subframe length must be consistent with the window function length, that is, completely aligned, and a frequency shift of a predetermined length is required to avoid spectrum leakage. The entire windowing process satisfies:

[0068] ;

[0069] The above formula is a formula representation of the windowing process of the lth frame.

[0070] in, represents the predicted speech signal of the lth frame after windowing, represents the lth frame predicted speech signal obtained by frame processing, l represents the frame index, H represents the frame shift, is a window function of fixed length.

[0071] Then perform a short-time Fourier transform to obtain:

[0072] ;

[0073] in, Represents the predicted speech signal of the first frame after windowing The short-time Fourier transform of represents the frequency index; N represents the Fourier transform length, i.e. the window length.

[0074] In the prior art, the parameter N in Fourier transform is generally a fixed parameter, such as 1024.

[0075] The invention introduces the idea of ​​fundamental frequency synchronization in the speech enhancement method of improving Wiener filtering based on harmonic model fundamental frequency synchronization, that is, the length of Fourier transform is automatically adjusted following the size of fundamental frequency, and the synchronization of fundamental frequency and short-time Fourier transform is achieved by introducing fundamental frequency parameters with semantic characteristics.

[0076] The predicted speech signal of the first frame after windowing The short-time Fourier transform Change to:

[0077] ;

[0078] in, It represents the window function of the lth frame, and the window function of each frame is changed from a fixed window length to a window length that changes flexibly with the base frequency.

[0079] ;

[0080] in, is the window length, i.e. the Fourier transform length, Follow the frame index l changes; is the sampling rate, is the fundamental frequency of the predicted speech signal of the lth frame obtained in step 2, and K is a constant.

[0081] Specifically, It follows Changing, And it changes with l.

[0082] Step 4. Perform harmonic modeling on the voiced frames in the predicted speech signal and obtain the reconstructed phase based on the harmonic model.

[0083] In the harmonic model, the speech signal is modeled as a superposition of sine waves whose amplitude, phase and frequency vary with time. Since the fundamental frequency of each speaker is different, the speech enhancement method of the present invention based on the fundamental frequency synchronization of the harmonic model improves the Wiener filter and takes the fundamental frequency as the starting point for harmonic modeling. The harmonic parameters including the number of harmonics and the Fourier transform length have a linear relationship with the fundamental frequency. Since the fundamental frequency of the voice of a bystander in the noise is different from that of the speaker, its speech characteristics will not be modeled, thereby achieving accurate filtering.

[0084] The time domain formula for harmonic modeling of the voiced frames in the predicted speech signal is expressed as:

[0085] ;

[0086] in and are the phase information and amplitude information of the rth harmonic respectively, Indicates the number of harmonics. The number of harmonics is determined by the sampling rate. and fundamental frequency Decide, satisfy:

[0087] ;

[0088] The frequency domain formula for harmonic modeling of the voiced frames in the predicted speech signal is expressed as:

[0089] ;

[0090] in , , are the fundamental frequency, phase, and amplitude of the lth frame of the rth harmonic respectively.

[0091] make: .

[0092] According to Euler's formula:

[0093] ;

[0094] get:

[0095] S l ( k ) = ∑ r = 1 R ( n ) A r , l 2 ∑ n = 0 M ( l ) − 1 [ w l ( n ) e − j 2 π n M ( l ) ( k − f r , l M ( l ) ) e j φ ′ r , l + w l ( n ) e − j 2 π n M ( l ) ( k + f r , l M ( l ) ) e − j φ ′ r , l ] ;

[0096] Using the frequency shift characteristics of Fourier transform, we get:

[0097] S l ( k ) = ∑ r = 1 R ( n ) A r , l 2 [ W l ( k − k ′ r , l ) M ( l ) e j φ ′ r , l + W l ( k + k ′ r , l ) M ( l ) e − j φ ′ r , l ] ;

[0098] in is the window function The Fourier transform form of is the frequency offset.

[0099] ;

[0100] Represents the window function The Fourier transform form of with a left frequency shift.

[0101] Represents the window function The Fourier transform form of with right frequency shift.

[0102] use The conjugate symmetry of , we get:

[0103] .

[0104] Converting to exponential form, we get:

[0105] .

[0106] in, yes phase information.

[0107] make ;

[0108] in, Represents the phase information of the predicted speech signal, and we get:

[0109] ;

[0110] when , That is, the reconstructed phase information of the lth frame obtained based on the fundamental frequency synchronization processing of the harmonic model.

[0111] ;

[0112] in, represents the phase information of the window function, is zero; therefore, the reconstructed phase information is obtained:

[0113] ;

[0114] Phase information of the l-1th frame It is expressed as:

[0115] ;

[0116] in, , represent the fundamental frequency and phase of the l-1th frame of the rth harmonic respectively.

[0117] Based on the short-term stationary characteristics, the fundamental frequency and phase changes between adjacent frames are the same, that is, there is a , ,therefore:

[0118] .

[0119] Reconstructing phase information It is expressed as:

[0120] .

[0121] Among them, ∠ is the symbol for phase information, represents the reconstructed phase information of frame 0, represents the reconstructed phase information of the lth frame, Represents the reconstructed phase information of the l-1th frame.

[0122] Step 5. Predict the speech signal output by the Wiener filter in step 1 The power spectrum The square root is used to obtain the amplitude spectrum of the enhanced speech signal, and then combined with the reconstructed phase information in step 4 , synthesize enhanced speech signal .

[0123] According to the power spectrum in step 1 Obtain the amplitude spectrum of the enhanced speech signal, and then combine it with the reconstructed phase information in step 4 , forming the enhanced speech signal , which is expressed as:

[0124] .

[0125] The amplitude spectrum of the enhanced speech signal in the present invention is obtained by using the Wiener filtering algorithm, and a harmonic model is constructed for fundamental frequency synchronization based on the linear relationship between the number of harmonics and the fundamental frequency. The phase information is reconstructed by the harmonic model in combination with fundamental frequency synchronization.

[0126] In addition, in order to verify the effectiveness of the method proposed in the present invention, the following specific experiments are also given:

[0127] Figure 2 and Figure 7 They are clean speech signals The time domain and spectrum diagram of Figure 3 and Figure 8 They are noisy speech signals The time domain and spectrum diagram of Figure 4 and Fig. 9They are noisy speech signals The time domain diagram and spectrum diagram of the output speech signal after spectral subtraction, Figure 5 and Fig.10 They are noisy speech signals The time domain diagram and spectrum diagram of the output speech signal after the Wiener filtering algorithm, Figure 6 and Fig.11 is a noisy speech signal The output speech signal obtained by the method of the present invention is the enhanced speech signal Time domain and spectrum diagram of .

[0128] It can be seen from the time domain diagram that Figure 4 The spectral subtraction noise filtering effect is the worst. Figure 6 The speech enhancement method of the present invention based on the harmonic model fundamental frequency synchronization improved Wiener filtering has the best noise filtering effect. Figure 5 The Wiener filtering algorithm has been significantly improved, and the waveform of the output speech signal is closest to the clean speech signal.

[0129] It can be seen from the spectrum diagram that Figure 7 Medium clean speech signal After noise processing, Figure 8 The noisy speech signal After that, the edge spectrum information of speech is masked. Fig. 9 Spectral subtraction and Fig.10 The Wiener filtering algorithms failed to restore the spectrum information normally. Fig.11 The output speech signal of the speech enhancement method of the present invention based on the harmonic model fundamental frequency synchronization improved Wiener filtering is Figure 7 The spectrum comparison of clean speech signals shows excellent similarity.

[0130] The above experiments show that the method proposed in the present invention is superior to other comparative methods.

[0131] Example 2

[0132] This embodiment 2 describes a speech enhancement system based on harmonic model fundamental frequency synchronization improved Wiener filtering, which is based on the same inventive concept as the speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering in embodiment 1.

[0133] Specifically, the speech enhancement system based on harmonic model fundamental frequency synchronization improved Wiener filtering includes the following modules:

[0134] The Wiener filter module is used to calculate the response function of the Wiener filter according to the minimum mean square error criterion, and obtain the power spectrum of the predicted speech signal output by the noisy speech signal after passing through the Wiener filter;

[0135] A fundamental frequency estimation module, used for obtaining fundamental frequency information of a predicted speech signal through a fundamental frequency estimation algorithm;

[0136] A Fourier transform module is used to perform windowing processing and baseband synchronization Fourier transform on the predicted speech signal to obtain a Fourier transform form of the predicted speech signal;

[0137] A phase reconstruction module, used for performing harmonic modeling on the voiced frames in the predicted speech signal, and obtaining reconstructed phase information based on the harmonic model;

[0138] And an enhanced speech synthesis module is used to obtain the amplitude spectrum of the enhanced speech signal by square rooting the power spectrum of the predicted speech signal output by the Wiener filter, and then synthesize the enhanced speech signal in combination with the reconstructed phase information of the phase reconstruction module.

[0139] It should be noted that in the speech enhancement system based on harmonic model fundamental frequency synchronization improved Wiener filtering, the implementation process of the functions and effects of each functional module is specifically described in the implementation process of the corresponding steps in the method in Example 1, and will not be repeated here.

[0140] Example 3

[0141] This embodiment 3 describes a computer device, which includes a memory and one or more processors.

[0142] The memory stores executable codes, and when the processor executes the executable codes, the executable codes are used to implement the steps of the speech enhancement method based on harmonic model fundamental frequency synchronization to improve Wiener filtering in the above-mentioned embodiment 1.

[0143] In this embodiment, the computer device is any device or apparatus with data processing capability, which will not be described in detail here.

[0144] Example 4

[0145] This embodiment 4 describes a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the steps of a speech enhancement method for improving Wiener filtering based on harmonic model fundamental frequency synchronization are performed.

[0146] The computer-readable storage medium may be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc., equipped on the device.

[0147] Of course, the above description is only a preferred embodiment of the present invention, and the present invention is not limited to the above embodiments. It should be noted that all equivalent substitutions and obvious deformation forms made by any technician familiar with the field under the guidance of this specification fall within the essential scope of this specification and should be protected by the present invention.

Claims

1. A speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering, characterized in that: The steps include: Step 1. Calculate the response function of the Wiener filter according to the minimum mean square error criterion to obtain the power spectrum of the predicted speech signal output by the noisy speech signal after passing through the Wiener filter; Step 2. Obtain the fundamental frequency information of the predicted speech signal through a fundamental frequency estimation algorithm; Step 3. Perform windowing processing and baseband synchronization Fourier transform on the predicted speech signal to obtain the Fourier transform form of the predicted speech signal; Step 4. Perform harmonic modeling on the voiced frames in the predicted speech signal, and obtain reconstructed phase information based on the harmonic model; Step 5. squaring the power spectrum of the predicted speech signal output by the Wiener filter in step 1 to obtain the amplitude spectrum of the enhanced speech signal, and then combining it with the reconstructed phase information in step 4 to synthesize the enhanced speech signal; The step 3 is specifically as follows: Based on the short-time stationary characteristics of speech signals and the requirements of short-time Fourier transform, the predicted speech signal is first Performing frame processing, and then performing windowing processing on the subframes obtained by the frame processing; The windowing process requires that the subframe length and the window function length are consistent, that is, completely aligned, and a frequency shift of a predetermined length is required to avoid spectrum leakage; the entire windowing process satisfies: in, represents the predicted speech signal of the lth frame after windowing, represents the lth frame predicted speech signal obtained by frame processing, w(n) is a window function of fixed length, and H represents the frame shift; Then perform a short-time Fourier transform to obtain: Among them, S l (k) represents the predicted speech signal of the first frame after windowing The short-time Fourier transform of , j represents the imaginary unit, k represents the frequency index; N represents the Fourier transform length, i.e., the window length; Introducing fundamental frequency synchronization in the Fourier transform process, that is, the Fourier transform length is automatically adjusted according to the fundamental frequency size, and the fundamental frequency and short-time Fourier transform are synchronized by introducing the fundamental frequency parameters with semantic features; The predicted speech signal of the first frame after windowing The short-time Fourier transform S l (k) is amended to read as follows: Among them, w l (n) represents the window function of the first frame, and its window length changes synchronously with the fundamental frequency; Where M(l) is the window length, i.e., the Fourier transform length, and M(l) changes with the frame index l; f s is the sampling rate, f 0,l Predict the fundamental frequency of the speech signal for the lth frame obtained in step 2, where K is a constant.

2. The speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering according to claim 1, characterized in that: The step 1 is specifically as follows: Assume that the noisy speech signal is x(n), the clean speech signal is s(n), and the noise signal is v(n). x(n) and s(n) are independent, and x(n), s(n) and v(n) satisfy: x(n)=s(n)+v(n); Where n is the time series of discrete speech signals; The predicted speech signal output by the noisy speech signal x(n) after passing through the Wiener filter is Then we have: Wherein, m represents the time index, p represents the order of the Wiener filter, x(nm) represents the noisy speech signal at time nm, and h(m) is the response function of the Wiener filter; The error signal ε satisfies: Mean square error E[(ε) 2 ]for: According to the minimum mean square error criterion, the partial derivative of h(m) is calculated and set to zero to obtain the covariance form: Among them, R sx (m) represents the cross-covariance function of the clean speech signal s(n) and the noisy speech signal x(n), l represents the frame index, h(l) represents the response function of the Wiener filter, R xx (ml) represents the autocovariance function of the noisy speech signal x(n); Then perform a short-time Fourier transform and get: Among them, H(e jw ) is the gain of the Wiener filter, P ss (e jw ) and P xx (e jw ) are the power spectra of the clean speech signal and the power spectra of the noisy speech signal, respectively, j represents the imaginary unit, and w represents the angular frequency; Predicting speech signals The power spectrum P is expressed as: P=H(e jw )·P xx (e jw )。 3. The speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering according to claim 1, characterized in that: In step 2, the fundamental frequency estimation algorithm adopts the fundamental frequency estimation filter PEFAC with amplitude compression.

4. The speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering according to claim 2, characterized in that: The step 4 is specifically as follows: In the harmonic model, the predicted speech signal is modeled as a superposition of sinusoidal waves whose amplitude, phase, and frequency vary over time; The time domain formula for harmonic modeling of the voiced frames in the predicted speech signal is expressed as: in and A r are the phase information and amplitude information of the rth harmonic respectively, R(n) represents the number of harmonics, and the number of harmonics is determined by the sampling rate f s Determined by the fundamental frequency f0, the number of harmonics R(n) satisfies: The frequency domain formula for harmonic modeling of the voiced frames in the predicted speech signal is expressed as: where f r,l , A r,l are the fundamental frequency, phase, and amplitude of the lth frame of the rth harmonic respectively; make: According to Euler's formula: get: Using the frequency shift characteristics of Fourier transform, we get: Where W l (k) is the window function w l The Fourier transform form of (n), k r ' ,l is the frequency offset; k r ' ,l =f r,l M(l); W l (kk r ' ,l ) M(l) Represents the window function w l The Fourier transform form of (n) with left frequency shift; W l (k+k r ' ,l ) M(l) Represents the window function w l (n) with right frequency shift; using S l The conjugate symmetry of (k) yields: Converting to exponential form, we get: Among them, φ l W (kk r ' ,l ) is W l (kk r ' ,l ) M(l) Phase information of make Among them, φ r,l (k) represents the phase information of the predicted speech signal, and we get: When k = k r ' ,l ,φ r,l (k r ' ,l ) is the reconstructed phase information of the first frame obtained based on the fundamental frequency synchronization processing of the harmonic model; Among them, φ l W (0) represents the phase information of the window function, φ l W (0) is zero, so the reconstructed phase information is obtained: The reconstructed phase information φ of the l-1th frame r,l-1 (k r ' ,l-1 ) is expressed as: Among them, f r,l-1 , denote the fundamental frequency and phase of the l-1th frame of the rth harmonic respectively; Based on the short-term stationary characteristics, the fundamental frequency and phase changes between adjacent frames are the same, that is, there is a constant f between adjacent frames. r,l-1 =f r,l , therefore: Reconstruct phase information ∠S l (k) is expressed as: Among them, ∠ is the symbol for phase information, ∠S l (k r ' ,0 ) represents the reconstructed phase information of frame 0, ∠S l (k r ' ,l ) represents the reconstructed phase information of the lth frame, ∠S l (k r ' ,l-1 ) represents the reconstructed phase information of the l-1th frame.

5. The speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering according to claim 4, characterized in that: The step 5 is specifically as follows: The predicted speech signal output by the Wiener filter according to the noisy speech signal in step 1 The power spectrum P of the enhanced speech is squared to obtain the amplitude spectrum, and then combined with the reconstructed phase information ∠S in step 4 l (k), synthesized enhanced speech signal Y l (k), which is expressed as:

6. A speech enhancement system based on harmonic model fundamental frequency synchronization improved Wiener filtering for implementing the speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering as claimed in claim 1, characterized in that: The speech enhancement system based on harmonic model fundamental frequency synchronization improved Wiener filtering includes: The Wiener filter module is used to calculate the response function of the Wiener filter according to the minimum mean square error criterion, and obtain the power spectrum of the predicted speech signal output by the noisy speech signal after passing through the Wiener filter; A fundamental frequency estimation module, used for obtaining fundamental frequency information of a predicted speech signal through a fundamental frequency estimation algorithm; A Fourier transform module is used to perform windowing processing and baseband synchronization Fourier transform on the predicted speech signal to obtain a Fourier transform form of the predicted speech signal; A phase reconstruction module, used for performing harmonic modeling on the voiced frames in the predicted speech signal, and obtaining reconstructed phase information based on the harmonic model; And an enhanced speech synthesis module is used to obtain the amplitude spectrum of the enhanced speech signal by square rooting the power spectrum of the predicted speech signal output by the Wiener filter, and then synthesize the enhanced speech signal in combination with the reconstructed phase information of the phase reconstruction module.

7. A computer device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, the steps of the speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering are implemented as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the speech enhancement method based on harmonic model fundamental frequency synchronization improved Wiener filtering as described in any one of claims 1 to 5 are implemented.

Images