Speech recognition method and system in complex noise background based on quadratic spectral subtraction

By using quadratic spectral subtraction in complex noise background, combining historical noise and noise estimation of current frame audio, the removal factor is calculated for speech signal processing, which solves the problem of noise estimation deviation under complex noise background and improves the accuracy of speech recognition and noise reduction effects.

CN115132219BActive Publication Date: 2024-11-19COMP APPL TECH INST OF CHINA NORTH IND GRP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210711617.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-22
Publication Date
2024-11-19
Estimated Expiration
2042-06-22

AI Technical Summary

Technical Problem

The prior art is difficult to effectively control residual noise under complex noise background, especially noise estimation deviation caused by noise background differences in different scenarios, affecting the speech recognition effect.

Method used

Using a method based on quadratic spectrum subtraction, the historical noise removal factor and the current frame audio noise removal factor are calculated through historical noise estimation and current frame audio noise estimation, and the speech signals under complex noise background are processed and the quadratic spectrum is performed. Reduce to reduce noise.

Benefits of technology

The accuracy of noise reduction processing of speech recognition is improved, and the noise reduction effect of the audio is significantly improved by iteratively optimizing historical noise estimation, reducing the error rate, and improving the accuracy of speech recognition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115132219B_ABST
    Figure CN115132219B_ABST
Patent Text Reader

Abstract

The present invention relates to a method and system for speech recognition in a complex noise background based on quadratic spectral subtraction, and belongs to the technical field of speech enhancement. The method comprises: selecting noisy historical audio and pure noise audio in a complex noise background, and obtaining a historical noise estimate through calculation and processing; performing frame processing on the audio to be recognized in the complex noise background to obtain multiple frames of audio; and processing each frame of audio in turn: determining a historical noise removal factor and a current frame audio noise removal factor based on the historical noise estimate and the noise estimate of the current frame audio, performing quadratic spectral subtraction on the current frame audio, and obtaining a speech spectrum after noise reduction of the current frame audio. The method solves the problem that the existing technology cannot control the residual noise to a low level for the complex background noise in the real world.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of speech enhancement, and in particular relates to a speech recognition method and system in a complex noise background based on quadratic spectral subtraction. Background Art

[0002] Spectral subtraction is one of the speech enhancement algorithms. Speech enhancement refers to the technology of extracting useful speech signals from noisy speech signals and suppressing and reducing noise interference when the speech signals are interfered with or even drowned by various noises. However, noise interference is usually random, and it is almost impossible to extract completely pure speech signals from noisy speech. In this case, there are two main purposes of speech enhancement: one is to improve speech quality and eliminate background noise, which is a subjective measure; the other is to improve the effects of speech tasks such as speech recognition and speaker recognition, which is an objective measure. However, these two purposes are often not achieved at the same time. For example, some speech enhancement algorithms can significantly reduce background noise and improve speech quality, but they cannot improve the effect of speech recognition, or even slightly reduce it.

[0003] Spectral subtraction is a traditional and effective speech enhancement algorithm for processing broadband noise. Its basic idea is to subtract the noise power spectrum from the noisy speech signal under the assumption that the additive noise and the short-term stable speech signal are independent of each other, so as to obtain a relatively pure speech spectrum. The outstanding advantages of spectral subtraction are that the algorithm has a relatively small amount of calculation, a relatively low amount of operation, a very low computational complexity, and is suitable for real-time processing scenarios. The disadvantage is that the processed signal will have a lot of residual noise, which is called musical noise.

[0004] In order to reduce the music noise caused by spectral subtraction, Berouti proposed a spectral over-subtraction algorithm, which uses a noise removal factor to reduce the amplitude of the broadband spectrum peak remaining after spectral subtraction, and uses the lowest audio energy to fill the spectrum valley (negative value of spectral subtraction) to control the amount of residual noise and the size of the music noise. The expression of over-subtraction spectral subtraction is as follows:

[0005]

[0006] Among them, P y (ω), P s (ω), P n (ω), respectively, represents the power spectra of the noisy signal, the clean speech signal and the noise signal; α is the noise removal factor, which is the coefficient of the audio spectrum minus the noise spectrum; b represents the minimum audio energy retained in the audio;

[0007] Spectral subtraction and over-subtraction spectral subtraction are both valid in a steady-state background noise environment, that is, the noise has the same degree of influence on all spectral components of speech. However, the background noise in the real world changes at any time, and different interference noises have different effects on different frequency bands of speech. Over-subtraction spectral subtraction still cannot control the residual noise to a low level. Summary of the invention

[0008] In view of the above analysis, the present invention aims to provide a speech recognition method in a complex noise background based on quadratic spectral subtraction, which performs noise estimation on noisy historical audio, estimates the noise of current audio, determines the historical noise removal factor and the current frame audio noise removal factor by calculation, and performs quadratic spectral subtraction on the noisy audio to be identified, so as to solve the problem that the background noise in the real world is complex and the residual noise cannot be controlled to a low level using existing technologies.

[0009] On the one hand, the present invention provides a method for speech recognition in a complex noise background based on quadratic spectral subtraction, which specifically comprises the following steps:

[0010] Based on noisy historical audio and pure noise audio under a complex noise background, obtaining a historical noise estimate of the complex noise background;

[0011] Perform frame processing on the audio to be recognized under the complex noise background to obtain multiple frames of audio;

[0012] Process each frame of audio in turn to obtain noise-reduced speech; wherein, processing the current frame of audio includes: performing secondary spectrum subtraction on the current frame of audio based on the historical noise estimation and the noise estimation of the current frame of audio to obtain the speech spectrum of the current frame of audio after noise reduction.

[0013] Furthermore, obtaining the historical noise estimation of the complex noise background based on the noisy historical audio and the pure noise audio under the complex noise background includes:

[0014] Each of the noisy historical audios is divided into frames, and after processing, a power spectrum of each frame signal of the noisy historical audio is obtained;

[0015] Select the preset number of frames of audio with the lowest power spectrum on each audio as pure noise, and estimate the average power spectrum B of each frame of the noise of each noisy historical audio. i (ω), where i = 1, 2, 3, ..., n, and n represents the number of noisy historical audios;

[0016] Each of the pure noise audio is divided into frames, and after processing, an average noise power spectrum B of each frame of each pure noise audio is obtained. j (ω), where j = 1, 2, 3, ..., k, k represents the number of pure noise audio;

[0017] B i (ω) and B j (ω) Take the average value to get the historical noise estimate

[0018] Furthermore, the noise estimation of the current frame audio includes:

[0019] Select a preset number of frames of audio with the lowest power spectrum in the audio to be recognized as pure noise, and estimate the average power spectrum of the noise of each frame of the audio to be recognized That is, the noise estimate of the current frame audio.

[0020] Further, the current frame audio is subjected to secondary spectrum subtraction using the following formula to obtain a speech spectrum after noise reduction of the current frame audio:

[0021]

[0022] in, Represents the power spectrum estimate of the current frame audio, Y n+1 (ω,m) represents the spectrum of the current frame audio, ψ n+1 (ω,m) represents the phase information of the current frame audio; α m , β m are the historical noise removal factor and the current frame audio noise removal factor respectively; b m It is the lowest spectral factor of the audio signal.

[0023] Further, the α is calculated using the following formula: m , β m and b m :

[0024]

[0025]

[0026]

[0027] Where c is a constant, ξ m is the a posteriori signal-to-noise ratio of the current frame audio signal in the frequency domain; α min , α max Respectively represent α m The minimum and maximum values ​​of β min , β max Respectively represent β m The minimum and maximum values ​​of b min 、b max Respectively represent b m The maximum and minimum values ​​of .

[0028] Further, the ξ is calculated using the following formula: m :

[0029]

[0030] Where k is the frequency point, ∑ k |Y n+1 (ω k , m)| represents the audio spectrum intensity of the current frame, represents the spectrum intensity of the historical noise estimate.

[0031] Furthermore, for the α m , β m and b m The maximum and minimum values ​​of α are restricted. max =3,α min =1,β max =3,β min =1,b max =0.1,b min =0.02.

[0032] On the other hand, the present invention also provides a speech recognition system in a complex noise background based on quadratic spectral subtraction, comprising:

[0033] Historical noise estimation module: based on noisy historical audio and pure noise audio under complex noise background, obtains historical noise estimation of the complex noise background;

[0034] The module for noise reduction processing of the audio to be identified is as follows: the audio to be identified in a complex noise background is processed by framing to obtain multiple frames of audio; each frame of audio is processed in turn to obtain the denoised speech; wherein, the processing of the current frame of audio includes: based on the historical noise estimation and the noise estimation of the current frame of audio, a secondary spectrum subtraction is performed on the current frame of audio to obtain the speech spectrum of the current frame of audio after denoising.

[0035] Furthermore, the historical noise estimation module includes the following modules:

[0036] Noisy historical audio processing module: each noisy historical audio is divided into frames, and the power spectrum of each frame signal of the noisy historical audio is obtained after processing; a preset number of frames of audio with the lowest power spectrum on each audio are selected as pure noise, and the average power spectrum B of each frame of the noise of each noisy historical audio is estimated. n (ω), where n represents the nth noisy historical audio;

[0037] Pure noise audio processing module: divide each pure noise audio into frames, and obtain the average noise power spectrum B of each frame of each pure noise audio after processing. k(ω), where k represents the kth pure noise audio;

[0038] Noise estimation module: B p (ω) and B n (ω) Take the average value to get the historical noise estimate

[0039] Furthermore, the to-be-recognized audio noise reduction processing module includes:

[0040] Current audio noise estimation module: Frame the audio to be identified under complex noise background to obtain multiple frames of audio, select a preset number of frames of audio with the lowest power spectrum as pure noise, and estimate the average power spectrum of the noise for each frame of the audio to be identified That is, the noise estimation of the current frame audio;

[0041] Noise reduction processing module: Based on the historical noise estimation and the noise estimation of the current frame audio, perform secondary spectrum subtraction on the current frame audio to obtain the speech spectrum of the current frame audio after noise reduction.

[0042] The present invention can achieve at least one of the following beneficial effects:

[0043] 1. By collecting a large number of historical noise frequencies of complex noise backgrounds in different scenes and performing noise-weighted averaging to obtain historical noise estimation, and by extracting specific frames of the current audio to be identified to perform current noise estimation, the problem of deviation between the noise estimation and the noise in the audio to be identified caused by the differences in complex noise backgrounds in different scenes is solved.

[0044] 2. By calculating and determining the historical noise removal factor and the current frame audio noise removal factor, a secondary spectrum subtraction is performed on the noisy audio to be identified, thereby improving the accuracy of the noise reduction processing on the audio to be identified.

[0045] 3. By iterating the noise estimation of the audio to be processed each time to the historical noise estimation, the historical noise estimation can be continuously optimized, and the accuracy of the noise reduction processing of the audio to be recognized by the quadratic spectral subtraction can be further improved.

[0046] Other features and advantages of the present invention will be described in the following description, and some advantages may become apparent from the description, or may be understood by practicing the present invention. The purpose and other advantages of the present invention may be realized and obtained through the contents particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0048] Figure 1 It is a schematic diagram of the process of the method of the present invention;

[0049] Figure 2 Detailed step diagram of the method of the present invention;

[0050] Figure 3 The figure is a waveform comparison of an audio to be processed before and after processing in one embodiment of the present invention. DETAILED DESCRIPTION

[0051] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0052] Method Embodiment

[0053] Example 1

[0054] A specific embodiment of the present invention discloses a method for speech recognition in a complex noise background based on quadratic spectral subtraction, which specifically includes the following steps:

[0055] Step S1: Select noisy historical audio and pure noise audio under complex noise background, and obtain historical noise estimation through calculation and processing.

[0056] Complex noise background refers to a speech signal environment in which there are various noises that are irrelevant to the presence or absence of the signal and the noise is constantly changing; noisy historical audio refers to audio that contains complex noise and pure speech signals, and pure noise audio refers to noise audio that does not contain speech signals captured in a complex noise background. This method can be applied to steady-state noise speech enhancement scenarios in the military field, such as speech recognition scenarios under the operation of armored vehicles; in this case, noisy historical audio refers to audio that contains noise generated by the operation of armored vehicles and pure speech signals, and pure noise audio refers to the noise generated when armored vehicles are running.

[0057] Specifically, step S1 includes the following steps:

[0058] Step S11: select i noisy historical audios (where i=1, 2, 3, ..., n, where n is the number of noisy historical audios), divide the i-th noisy audio into frames, illustratively, divide one frame every 25 ms, and shift the frame by 10 ms, and use x(i,m), y(i,m) and b(i,m) to represent the clean audio, noisy audio and noisy audio signals at the m-th frame in the i-th audio, respectively, and the corresponding spectra are represented by X(ω,m), Y(ω,m) and B(ω,m), respectively;

[0059] Step S12: Perform short-time Fourier transform on each frame of audio to obtain the power spectrum X of each frame signal. i(ω,m) and phase spectrum ψ i (ω,m);

[0060] Step S13: Select the power spectrum X of the entire audio i The lowest 100 frames of audio (ω, m) are taken as pure noise, and the average power spectrum B of each frame of the noise to be reduced is estimated. i (ω);

[0061] Step S14: Select j pure noise audios (where j = (1, 2, 3, ..., k)), divide the jth pure noise audio into frames, divide it into frames every 25 ms, shift it by 10 ms, perform short-time Fourier transform on each frame of audio and calculate the weighted power spectrum to obtain the average noise power spectrum B of each frame j (ω);

[0062] Step S15: B i (ω) and B j (ω) Take the average value to get the historical noise estimate

[0063] Step S2: dividing the audio to be recognized under the complex noise background into frames, and processing each frame of the audio in turn to obtain the noise-reduced speech, which specifically includes the following steps:

[0064] Step S21: Let the sequence number of the audio to be identified be n+1, let i=n+1, and execute the same steps as S11, S12, and S13 to estimate the average power spectrum of each frame of the audio to be identified.

[0065] Step S22: Calculate the parameters required for performing secondary spectral subtraction on the current frame audio: m , α m , β m and b m The value of ξ m is the posterior signal-to-noise ratio of the current frame audio signal frequency domain; α m , β m are the historical noise removal factor and the current frame audio noise removal factor respectively; b m is the lowest spectral factor of the audio signal;

[0066]

[0067] Where k is the frequency point, ∑ k |Y n+1 (ω k , m)| represents the audio spectrum intensity of the current frame, A spectrum intensity representing the historical noise estimate;

[0068]

[0069]

[0070]

[0071] Step S23: Perform secondary spectrum subtraction on the current frame of the audio to be recognized using the following formula to obtain the speech spectrum after the current frame of audio noise reduction:

[0072]

[0073] It should be noted that, in the next recognition process of the audio to be recognized, the current audio to be recognized is added to the historical noisy audio as the n+1th audio, and the noise estimation of the audio to be recognized is Iterate to historical noise estimate Optimize historical noise estimates.

[0074] Example 2

[0075] Another specific embodiment of the present invention discloses a method for speech recognition in a complex noise background based on quadratic spectral subtraction, which specifically includes the following steps:

[0076] Step S4: 100 noisy human voice audio samples with a sampling rate of 44kHz and a length of 1 minute sampled at an armored vehicle training ground are selected as noisy historical audio; wherein the background noise scenes include: only the engine running noise when the armored vehicle is stationary, the engine running noise when the armored vehicle accelerates, the collision noise between the track and the wheel or the ground when the vehicle is moving, and the vehicle running noise when the armored vehicle decelerates, etc.; 300 armored vehicle noise samples with a sampling rate of 44kHz and a length of 1 minute sampled at the training ground are selected as pure noise audio; the historical noise estimation is obtained through calculation and processing, which specifically includes the following steps:

[0077] Step S41: for each noisy frequency, divide it into frames every 25 ms, and shift the frames by 10 ms. The clean frequency, noisy frequency and noisy audio signal at the mth frame in the i-th (i=1, 2, 3..., 100) audio are represented by x(i, m), y(i, m) and b(i, m), respectively. The corresponding spectra are represented by X(ω, m), Y(ω, m) and B(ω, m), respectively.

[0078] Step S42: Perform short-time Fourier transform on each frame of audio to obtain the power spectrum X of each frame signal. i (ω, m) and phase spectrum ψ i (ω, m);

[0079] Step S43: Select the power spectrum X of the entire audio iThe lowest 100 frames of audio (ω, m) are taken as pure noise, and the average power spectrum B of each frame of the noise of the audio to be reduced is estimated. i (ω);

[0080] Step S44: Divide each pure noise audio into frames, with one frame every 25 ms and a frame shift of 10 ms. Perform short-time Fourier transform on each frame of audio and calculate the weighted power spectrum to obtain the average noise power spectrum B of each frame. j (ω)(j=1, 2, 3..., 80);

[0081] Step S45: B i (ω) and B j (ω) Take the average value to get the historical noise estimate

[0082] Step S5: 50 noisy human voice audio samples with a sampling rate of 44kHz and a length of 1 minute from an armored vehicle training ground are selected as the audio to be recognized, the audio to be recognized is divided into frames, and each frame of audio is processed in turn to obtain the denoised speech, which specifically includes the following steps:

[0083] Step S51: Let the sequence number of the current audio to be recognized be n+1; let i=n+1, execute steps S41, S42, S43, and estimate the average power spectrum of each frame of the audio to be recognized

[0084] Step S52: Calculate the parameters required for performing secondary spectral subtraction on the current frame audio: m , α m , β m and b m The value of ξ m is the posterior signal-to-noise ratio of the current frame audio signal frequency domain; α m , β m are the historical noise removal factor and the current frame audio noise removal factor respectively; b m is the lowest spectral factor of the audio signal;

[0085]

[0086] Where k is the frequency point, ∑ k |Y n+1 (ω k , m)| represents the audio spectrum intensity of the current frame, A spectrum intensity representing the historical noise estimate;

[0087]

[0088]

[0089]

[0090] Step S53: Perform secondary spectrum subtraction on the current frame of the audio to be recognized using the following formula to obtain the speech spectrum after the current frame of audio noise reduction:

[0091]

[0092] Step S54: Estimating the noise of the current audio to be recognized Iterate to historical noise estimate Optimizing historical noise estimates;

[0093] Step S55: Execute steps S51, S52, S53, and S54 for the next audio to be recognized, until 50 audios to be recognized are processed.

[0094] The following method is used to verify the noise reduction effect of this embodiment:

[0095] Use the annotation tool to annotate the speaking voices in the 50 audio files to be recognized. The time slot is 10ms. For each time slot, if there is a human voice, it is marked as 1, otherwise it is marked as 0; each audio file corresponds to a annotation file.

[0096] For each audio segment, since the audio length is fixed, the result after noise reduction processing is the same as the length of the annotated data. Compare the result data with the annotated data content, count the number of different annotated contents a, and calculate with the total audio annotation length b, and you can get the error rate e = a / b.

[0097] After calculation, the error rate of denoising the audio to be recognized under a complex noise background using a speech recognition method based on quadratic spectral subtraction disclosed in this embodiment is 14.92%, and the accuracy rate is 85.1%; the error rate of denoising 50 audios to be recognized using a unified power-based spectral subtraction method is 19.10%, and the accuracy rate is 80.9%; the accuracy of the speech recognition method of this embodiment is significantly improved compared with the classical spectral subtraction method.

[0098] Figure 3 A waveform comparison of an audio to be processed before and after being processed using the speech recognition method of this embodiment is given. The waveforms before and after the sound noise reduction can be observed from the marked positions in the figure, and the background noise removal effect is obvious.

[0099] System Example

[0100] A speech recognition system in a complex noise background based on quadratic spectral subtraction, comprising:

[0101] Historical noise estimation module: based on noisy historical audio and pure noise audio under complex noise background, obtains historical noise estimation of the complex noise background;

[0102] The module for noise reduction processing of the audio to be identified is as follows: the audio to be identified in a complex noise background is processed by framing to obtain multiple frames of audio; each frame of audio is processed in turn to obtain the denoised speech; wherein, the processing of the current frame of audio includes: based on the historical noise estimation and the noise estimation of the current frame of audio, a secondary spectrum subtraction is performed on the current frame of audio to obtain the speech spectrum of the current frame of audio after denoising.

[0103] Among them, the historical noise estimation module includes the following modules:

[0104] Noisy historical audio processing module: each noisy historical audio is divided into frames, and the power spectrum of each frame signal of the noisy historical audio is obtained after processing; a preset number of frames of audio with the lowest power spectrum on each audio are selected as pure noise, and the average power spectrum B of each frame of the noise of each noisy historical audio is estimated. n (ω), where n represents the nth noisy historical audio;

[0105] Pure noise audio processing module: divide each pure noise audio into frames, and obtain the average noise power spectrum B of each frame of each pure noise audio after processing. k (ω), where k represents the kth pure noise audio;

[0106] Noise estimation module: B p (ω) and B n (ω) Weighted average to get historical noise estimate

[0107] The audio noise reduction processing module to be identified includes:

[0108] Current audio noise estimation module: Frame the audio to be identified under complex noise background to obtain multiple frames of audio, select a preset number of frames of audio with the lowest power spectrum as pure noise, and estimate the average power spectrum of the noise for each frame of the audio to be identified That is, the noise estimation of the current frame audio;

[0109] Noise reduction processing module: Based on the historical noise estimation and the noise estimation of the current frame audio, perform secondary spectrum subtraction on the current frame audio to obtain the speech spectrum of the current frame audio after noise reduction.

[0110] Since the speech recognition system in a complex noise background based on quadratic spectral subtraction and the speech recognition method in a complex noise background based on quadratic spectral subtraction are based on the same inventive concept, the relevant parts can be borrowed from each other, and thus the same technical effects can be achieved.

[0111] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A speech recognition method in a complex noise background based on quadratic spectral subtraction, characterized in that: The steps include: Based on noisy historical audio and pure noise audio under a complex noise background, an estimation of the historical noise of the complex noise background is obtained, including: dividing each of the noisy historical audio into frames, and obtaining a power spectrum of each frame signal of the noisy historical audio after processing; selecting a preset number of frames of audio with the lowest power spectrum on each audio as pure noise, and estimating an average power spectrum B of each frame of the noise of each noisy historical audio i (ω), where i = 1, 2, 3, ..., n, where n represents the number of noisy historical audios; each of the pure noise audios is divided into frames, and after processing, an average noise power spectrum B of each frame of each pure noise audio is obtained. j (ω), where j = 1, 2, 3, ..., k, where k represents the number of pure noise audio. i (ω) and B j (ω) takes the average value to obtain the historical noise estimate Perform frame processing on the audio to be recognized under the complex noise background to obtain multiple frames of audio; Process each frame of audio in turn to obtain a de-noised speech; wherein, processing the current frame of audio includes: performing secondary spectral subtraction on the current frame of audio based on the historical noise estimation and the noise estimation of the current frame of audio to obtain a speech spectrum of the current frame of audio after de-noising; wherein, the noise estimation of the current frame of audio includes selecting a preset number of frames of audio with the lowest power spectrum in the audio to be identified as pure noise; estimating the average power spectrum of the noise of each frame of the audio to be identified based on the pure noise That is, the noise estimation of the current frame audio; the current frame audio is subjected to secondary spectrum subtraction using the following formula to obtain the power spectrum estimation of the current frame audio, that is, the speech spectrum of the current frame audio after noise reduction: in, represents the power spectrum estimation of the current frame audio, m represents the sequence number of the current frame audio, Y n+1 (ω, m) represents the spectrum of the current frame audio, ψ n+1 (ω, m) represents the phase information of the current frame audio; α m , β m are the historical noise removal factor and the current frame audio noise removal factor respectively; b m is the lowest spectral factor of the audio signal; the α is calculated using the following formula m , β m and b m : Where c is a constant, ξ m is the a posteriori signal-to-noise ratio of the current frame audio signal in the frequency domain; α min , α max Respectively represent α m The minimum and maximum values ​​of β min , β max Respectively represent β m The minimum and maximum values ​​of b min , b max Respectively represent b m The maximum and minimum values ​​of .

2. The speech recognition method according to claim 1, characterized in that: The ξ is calculated using the following formula: m : Where k is the frequency point, ∑ k |Y n+1 (ω k , m)| represents the audio spectrum intensity of the current frame, represents the spectrum intensity of the historical noise estimate.

3. The speech recognition method according to claim 2, characterized in that: For the α m , β m and b m The maximum and minimum values ​​of a max =3,a min =1,β max =3,β min =1,b max =0.1,b min = 0.

02.

4. A speech recognition system in a complex noise background based on quadratic spectral subtraction, characterized in that: include: Historical noise estimation module: based on noisy historical audio and pure noise audio under complex noise background, obtains historical noise estimation of the complex noise background; The historical noise estimation module includes: a noisy historical audio processing module, which is used to frame each of the noisy historical audios, and obtain the power spectrum of each frame signal of the noisy historical audio after processing; select a preset number of frames of audio with the lowest power spectrum on each audio as pure noise, and estimate the average power spectrum B of each frame of the noise of each noisy historical audio i (ω), where i = 1, 2, 3, ..., n, where n represents the number of noisy historical audios; a pure noise audio processing module is used to frame each of the pure noise audios, and after processing, obtain an average noise power spectrum B of each frame of each pure noise audio. j (ω), where j = 1, 2, 3, ..., k, where k represents the number of pure noise audio. i (ω) and B j (ω) takes the average value to obtain the historical noise estimate The module for noise reduction processing of the audio to be identified is as follows: the audio to be identified in the complex noise background is processed by framing to obtain multiple frames of audio; each frame of audio is processed in turn to obtain the speech after noise reduction; wherein, the processing of the current frame of audio includes: based on the historical noise estimation and the noise estimation of the current frame of audio, the secondary spectrum subtraction is performed on the current frame of audio to obtain the speech spectrum after noise reduction of the current frame of audio; The to-be-recognized audio noise reduction processing module includes: a current audio noise estimation module, which is used to treat a preset number of frames of audio with the lowest power spectrum in the to-be-recognized audio as pure noise; and to estimate the average power spectrum of noise of each frame of the to-be-recognized audio based on the pure noise. That is, the noise estimation of the current frame audio; a noise reduction processing module, which is used to perform secondary spectrum subtraction on the current frame audio based on the historical noise estimation and the noise estimation of the current frame audio to obtain the speech spectrum of the current frame audio after noise reduction; and perform secondary spectrum subtraction on the current frame audio using the following formula to obtain the power spectrum estimation of the current frame audio, that is, the speech spectrum of the current frame audio after noise reduction: in, represents the power spectrum estimation of the current frame audio, m represents the sequence number of the current frame audio, Y n+1 (ω, m) represents the spectrum of the current frame audio, ψ n+1 (ω, m) represents the phase information of the current frame audio; α m , β m They are the historical noise removal factor and the current frame audio noise removal factor; b m is the lowest spectral factor of the audio signal; the α is calculated using the following formula m , β m and b m : Where c is a constant, ξ m is the a posteriori signal-to-noise ratio of the current frame audio signal in the frequency domain; α min , α max Respectively represent α m The minimum and maximum values ​​of β min , β max Respectively represent β m The minimum and maximum values ​​of b min 、b max Respectively represent b m The maximum and minimum values ​​of .