CN102982801B - Phonetic feature extracting method for robust voice recognition - Google Patents

Abstract

The invention discloses a speech feature extraction method for robust speech recognition. The method comprises: obtaining a power spectrum; processing the power spectrum by using a filter bank; obtaining a medium-time power spectrum by means of frame averaging; The power spectrum is processed by asymmetric filtering, and the power spectrum is masked at the same time to obtain the pure speech power spectrum; the ratio of the pure speech and noisy speech power spectrum is subjected to channel average processing for smoothing; the smoothed pure speech and the banded speech The power spectrum ratio of noisy speech is multiplied by the power spectrum output by the filter bank to obtain the short-term power spectrum of pure speech; energy normalization is performed on the short-term power spectrum to eliminate multiplicative noise; the power spectrum is increased by equal loudness ; Exponential operation of the power spectrum; inverse Fourier transform of the power spectrum; calculation of the cepstral coefficient of the signal; mean normalization of the cepstral coefficient. The feature of the speech signal extracted by the present invention is fast and can be processed online; the acoustic model trained by using the feature extracted by the present invention has a good anti-noise effect; the present invention has very important application significance.

Description

A Speech Feature Extraction Method for Robust Speech Recognition

technical field

The invention relates to the field of speech recognition, in particular to a speech feature extraction algorithm method capable of significantly suppressing stationary and non-stationary noises in speech recognition.

Background technique

It is one of the most important problems in speech recognition that the recognition performance of the speech recognition system degrades sharply in complex environments. For example, when querying geographical location by voice on the mobile phone on the road, the acoustic environment in which the user is located is very complex and changes rapidly, which has a great impact on the performance of the voice recognition system. The original speech recognition system can achieve good processing and recognition results in a noise-free environment, but in practical applications, the performance of the recognition system will be affected by time-varying unpredictable environmental noise and channels, and speaker differences , The impact of factors such as changes in the content of the conversation has declined severely. So how to improve the robustness of the speech recognition system under the condition of mismatch between training and testing environment becomes the key of speech recognition technology.

In recent years, people have proposed many improved technologies and algorithms in the research field of environmental robustness of speech recognition technology, and achieved certain results. According to the process of speech recognition, robust speech recognition can be divided into four categories: anti-noise in the time-frequency domain; noise compensation in the feature domain; noise adaptation in the model domain and adaptation in the decoding domain. The earliest technology is anti-noise in the time-frequency domain, such as spectral subtraction and Wiener filtering, and the classic two-stage Wiener filtering of the European Telecommunications Standards Institute. Feature-level noise suppression usually compensates for noise during feature extraction. Since the PLP and MFCC features have always been at the top, the noise suppression at the feature level is mostly performed on these two features, such as vector Taylor series. The third stage is to adapt the noise in terms of models, including multi-state speech models, HMMs with shared variable parameters, etc. The fourth level is noise adaptation at the decoding level, including uncertainty decoding and sub-band reestimation instead of uncertainty decoding.

All of these methods are fundamentally seeking an optimal compensation for the mismatch between the training environment and the test environment under some criterion. Under a series of assumptions, such as Gaussian distribution of additive noise, independence of noise and speech signal, independence between different noises, gradient characteristics of channels, etc., these methods are made for the robustness of speech recognition technology. It has made useful exploration and contribution, especially in the stable noise environment, it has a better noise suppression effect. But there is still a big gap between this and the application requirements of the speech recognition system in the real noise environment, so it is powerless for more complex environments, such as sudden noise and other environments.

Contents of the invention

(1) Technical problems to be solved

In order to solve the above-mentioned shortcoming that the speech recognition rate in a complex environment is low, and the common feature extraction method is not strong enough to suppress non-stationary noise, the present invention proposes a feature extraction method that can improve its recognition rate, and the purpose is to improve The recognition rate of speech affected by additive noise such as burst noise and musical noise, and the speech recognition rate in a pure environment will not decrease.

(2) Technical solution

A kind of speech feature extraction method for robust speech recognition based on the present invention, comprises the following steps to realize:

Step 1, obtaining the power spectrum of the speech signal;

Step 2, processing the obtained power spectrum through a filter bank to obtain a short-term power spectrum of noisy speech;

Step 3, according to the obtained short-term power spectrum of the noisy speech, adopt the frame average mode to obtain the medium-duration power spectrum of the noisy speech;

Step 4, performing asymmetric filtering and masking anti-noise on the obtained medium-duration power spectrum of the noisy speech to obtain the medium-duration power spectrum of the pure speech;

Step 5, obtaining the short-term power spectrum of the pure speech according to the medium-duration power spectrum of the pure speech, the medium-duration power spectrum of the noisy speech, and the short-term power spectrum of the noisy speech;

Step 6, performing energy normalization processing on the short-term power spectrum of the pure speech to eliminate multiplicative noise;

Step 7, carrying out equal loudness emphasis on the short-term power spectrum of the pure speech that has eliminated the multiplicative noise;

Step 8, performing an exponential nonlinear operation on the short-term power spectrum of the pure speech after equal loudness emphasis;

Step 9. Inverse Fourier transform is performed on the short-term power spectrum of the pure speech after the exponential nonlinear operation to obtain cepstral coefficients, and the cepstral coefficients are subjected to mean value normalization processing to finally obtain speech features.

The present invention starts with the traditional speech feature extraction method, and aims at the disadvantage of weak anti-noise ability of the traditional speech feature, proposes several means to improve the speech feature, and finally forms a set of new speech feature extraction methods. Aiming at the characteristic that noise changes slower than speech, the present invention converts short-time power spectrum into medium-time power spectrum by means of frame average for estimating noise; uses asymmetric filtering to estimate noise and speech in noisy speech respectively. The spectral envelope of the spectrum; on the basis of asymmetric filtering, the signal-to-noise ratio is estimated by masking, and it is processed, and it is converted into the signal-to-noise ratio of the short-term power spectrum for anti-noise; also by energy normalization and Exponential nonlinearity processes the power spectrum. The speech feature extraction method for robust speech recognition proposed by the present invention can not only estimate the noise more accurately, but also make the speech features more in line with the auditory characteristics of human ears. Therefore, the features obtained by this feature extraction method have a good suppression effect on noise.

(3) Beneficial effects

The present invention starts with the traditional speech feature extraction method, and adds anti-noise processing and transformation processing in line with human hearing in the traditional speech feature extraction method, so that this feature extraction method can not only simulate various additive noises, but also The recognition rate in a pure environment is also higher than that of traditional speech feature extraction methods.

Description of drawings

Fig. 1 is the general flowchart of the speech feature extraction method that the present invention is used for robust speech recognition;

Fig. 2 is the structure flowchart that comprises the asymmetric low-pass filtering anti-noise module of masking;

FIG. 3 is a structural flowchart of the masking module in FIG. 2 .

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be described in further detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

FIG. 1 is a block diagram of the general flow of the speech feature extraction method for robust speech recognition in the present invention. As shown in Figure 1, a speech feature extraction method for robust speech recognition proposed by the present invention is mainly composed of the following processes: pre-emphasizing the speech signal; windowing the speech and using short-time Fourier transform to obtain Speech spectrum; square the speech spectrum to obtain the power spectrum; use the filter bank to process the power spectrum to obtain the short-term power spectrum of the noisy speech; use frame averaging to obtain the medium-duration power spectrum of the noisy speech; Perform asymmetric low-pass filtering on the medium-duration power spectrum of the noisy speech to track the noise in the speech, and at the same time mask the obtained medium-duration power spectrum to obtain the short-term power spectrum of pure speech; Perform channel average processing on the ratio of the power spectrum of the pure speech and the noisy speech for smoothing; compare the power spectrum ratio of the smoothed pure speech and the noisy speech with the short-term power spectrum of the noisy speech output by the filter bank Multiply the short-term power spectrum of pure speech to obtain the short-term power spectrum of pure speech; perform energy normalization processing on the short-term power spectrum of pure speech to eliminate multiplicative noise; perform equal loudness emphasis on the normalized short-term power spectrum to make it conform to Human auditory effect; then the power spectrum after equal loudness aggravation is converted into intensity and loudness by exponential operation, so that it conforms to human physiological characteristics; after that, inverse Fourier transform is performed on the power spectrum after intensity and loudness conversion; then according to the inverse Fourier transform The cepstral coefficients are calculated from the obtained results; finally, the mean value normalization process is performed on the obtained cepstral coefficients, and finally the speech features of the method of the present invention are obtained. Each step of the invention is described in detail below.

1. Pre-emphasize the voice signal

The purpose of pre-emphasis is to weaken the influence of low-frequency interference and highlight the main components of high-frequency signals. Usually, the following formula is used to pre-emphasize the speech sampling points:

${\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}& \mathrm{<span\; class="notranslate">\; if</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ <span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}& \mathrm{<span\; class="notranslate">\; if</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, α is called the pre-emphasis coefficient, x is the voice sampling point, y is the value of the voice sampling point after pre-emphasis, and t is the index of the sampling point.

Two, windowing the speech signal after the pre-emphasis and adopting short-time Fourier transform to obtain the speech spectrum Inside is a steady-state signal, and this segment of speech is called a frame. In order to reduce the truncation effect, a window is usually added to this speech, and the common ones are Hanning window and Hamming window. The frequency spectrum of this frame of speech can be obtained by performing short-time Fourier transform on a frame of speech signal after windowing. Specifically include: framing the voice, where the frame length range is 20ms to 30ms, and the frame shift range is 10ms to 15ms; adding a window to each frame of the voice, using a Hanning window or a Hamming window; Short-time Fourier transform is performed on the final speech, and the original Fourier transform formula is used, or the speech spectrum is obtained by filling the windowed speech with zeros to the integer power of 2 by fast Fourier transform.

3. Find the power spectrum by squaring the speech spectrum

In order to obtain the power spectrum P(w) of the speech signal, we square and sum the real and imaginary parts after the short-time Fourier transform respectively. The formula looks like this:

P(w)＝Re[S(w)] ² +Im[S(w)] ² (2)

where S(w) represents the short-time Fourier spectrum, and Re[S(w)] and Im[S(w)] represent the real and imaginary parts of the short-time Fourier spectrum, respectively.

Fourth, the use of filter banks to process the power spectrum

The human ear has different perceptual abilities to different frequencies of speech. Experiments have found that below 1000 Hz, the perceptual ability has a linear relationship with the frequency, while above 1000 Hz, the perceptual ability has a logarithmic relationship with the frequency. In order to simulate the human ear's perceptual characteristics of speech at different frequencies, a filter bank is usually used to transform the linear spectrum. The filter bank used can be a Mel-filter bank or a Gamma-tone filter bank, and the number of channels can be selected according to different filters;

In a preferred embodiment of the invention a Gamma-tone filter bank is used. It has several channels whose center frequencies are distributed linearly in the equivalent rectangular bandwidth.

Then the short-term power spectrum of the noisy speech is obtained by summing the Gamma-tone filter bank as follows:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}<span\; class="notranslate">\; ]</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{{\mathrm{<span\; class="notranslate">\; jw</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

Among them, m and l represent the index of frame and channel respectively, K is the point number of Fourier transform, w _k =2π/F _s , and F _s represents the sampling frequency of the speech signal. Indicates the mth frame of speech the magnitude of the frequency, then it means the lth channel Frequency gamma-tone filter value.

5. Calculate the medium-duration power spectrum by means of frame averaging

Since the change of noise is often slower than the change of speech, it is necessary to obtain a longer window than the ordinary window when estimating the noise. In the feature extraction method of the present invention, the average value of several windows is obtained by frame averaging to describe A longer window. But it is not possible to use such a long window for all voices, because if the window length is too large, the voice recognition rate will be reduced. The formula for calculating the medium-duration power spectrum of noisy speech by frame averaging is expressed as follows:

$\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\underset{{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Among them, m and l represent the index of frame and channel respectively, and M represents the sum of the number of frames taken forward and backward when calculating the medium duration.

6. Asymmetric filtering and masking anti-noise for medium-duration power spectrum of noisy speech

Since the noise changes rapidly at certain frequencies, in order to track the noise more accurately, the noise of different channels needs to be processed differently, so an asymmetric low-pass filter anti-noise module including masking is introduced here. The specific process is shown in Figure 2.

Fig. 2 is a structural flowchart of the masked asymmetric low-pass filter anti-noise module in the present invention.

In Figure 2, the first asymmetric low-pass filter can be described by the following formula:

${\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; let\text{'}s\; go</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; let\text{'}s\; go</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if\; Q</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&Greater\; Equal;</span>{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; let\text{'}s\; go</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>\\ {\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; b</span>}}{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; let\text{'}s\; go</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if\; Q</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; let\text{'}s\; go</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Among them, λ _a and λ _b are adjustable parameters, and the value range is (0-1). Q _le [m, l] obtained by the above formula is subtracted from Q [m, l], and then Q _o [m, l] is obtained through a half-wave rectification module. The specific operations of subtraction and rectification are shown in formula (6).

Q _o [m, l] = max(Q [m, l] - Q _le [m, l], 0) (6)

Q _o [m, l] are respectively fed into the masking module and the second asymmetric low-pass filter. The second asymmetric low-pass filter is the same as the above-mentioned first asymmetric low-pass filter, and the internal structure can still be obtained by formula (5), but the input is changed from Q[m, l] to Q _o [m , l], the output changes from Q _le [m, l] to Q _f [m, l]. The value Q _f [m, l] produced by the second asymmetric low-pass filter will be used as the spectral floor power, that is, the minimum value of the power spectrum. The second asymmetric low-pass filter is to prevent the output value of asymmetric filtering and masking the overall effect from being too small and causing unnecessary musical noise. On the other hand, Q _o [m, l] obtains Q _tm [m, l] through the masking module, and this step will be described in detail later. Q _tm [m, l] and Q _f [m, l] are jointly input to the maximum value module, and R _sp [m, l] is obtained by using the following formula:

R _sp [m, l] = max(Q _tm [m, l], Q _f [m, l]) (7)

Finally, a selective switch is used to determine the value of the result output R[m, l]. This optional switch is determined by the following formula:

$\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; sp</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; if\; Q</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&Greater\; Equal;</span>{\mathrm{<span\; class="notranslate">\; cQ</span>}}_{\mathrm{<span\; class="notranslate">\; let\text{'}s\; go</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>\\ {\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; if\; Q</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; cQ</span>}}_{\mathrm{<span\; class="notranslate">\; let\text{'}s\; go</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

Where c is an adjustable parameter, for example, c=2 can be selected. The significance of this formula is that if the medium duration power of a speech segment cannot be greater than its own spectral bottom power of c=2 times, it is considered that this section of speech is silent, so the output value should be the spectral bottom power.

The above-mentioned processing procedures are all realized by the asymmetric low-pass filter anti-noise module including masking described in FIG. 2 . The masking module in Fig. 2 is described in detail below. Its structure is shown in Figure 3. First input Q _o [m, l] to get Q _p [m, l] through the MAX module, the formula is as follows:

Q _p [m, l] = max(λ _t Q _p [m-1, l], Q _o [m, l]) (9)

Among them, λ _t is the forgetting coefficient, and the value range is (0~1). The final output value Q _tm [m, l] of the masking module is also determined by the selective switch, and its formula is described as follows:

${\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; tm</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>& {\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&Greater\; Equal;</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>\\ {\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>& {\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; )</span>$

Where μ _t is the corresponding parameter, and the value range is (0-1). The output value Q _tm [m, l] of the masking module is the same as the output Q _f [m, l] of the second asymmetric filter. After the maximum module described in the publicity (7), R sp [m, l] is obtained, and R _sp [m, l], R _sp [m, l] is finally the same as the spectral bottom Q _f [m, l] after the selective switch described by formula (8), and finally obtain the medium-duration power spectrum of noisy speech after asymmetric filtering and masking anti-noise Result R[m,l].

Through the above description, the output R[m, l] after asymmetric filtering and masking anti-noise is calculated. This value represents the medium-duration power spectrum of pure speech, which is the same as the medium-duration power spectrum Q of noisy speech after frame averaging. The ratio of [m, l] can describe the proportion of pure speech power in the noisy speech power spectrum, and we use H[m, l] to represent it. The formula is as follows:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>}{\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 11</span>}<span\; class="notranslate">\; )</span>$

7. Channel averaging and anti-noise integration

Since the thresholds are different from channel to channel, and processing is often based on a speech segment, smoothing between channels is necessary. We use the following formula for channel averaging to obtain the channel average weight H _s [m, l]:

${\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\underset{{\mathrm{<span\; class="notranslate">\; l</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}{\overset{{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 12</span>}<span\; class="notranslate">\; )</span>$

Wherein l ₂ =min(l+N, L), l ₁ =max(lN, 1), L represents the number of filter channels, N represents the total number of channels looking forward and backward when calculating the channel average . The channel-averaged weight H _s [m, l] is used to modulate the short-term power spectrum of noisy speech to obtain the short-term power spectrum of pure speech, the formula is as follows:

T[m,l]=P[m,l]H _s [m,l] (13)

8. Perform energy normalization on the short-term power spectrum of pure speech to eliminate multiplicative noise

In traditional feature extraction algorithms such as MFCC, in order to fit the physiological characteristics of people, logarithmic operations are used, so that the noise caused by the multiplicative operation in the feature extraction algorithm is changed into additive information, and finally through the mean Normalization removed. However, in the feature extraction method of the present invention, exponential operations are used to fit human physiological characteristics, so that the noise caused by multiplicative operations cannot be eliminated by mean value normalization, so this step is added in order to be able to Remove this multiplicative noise.

Since the feature extraction method of the present invention is an online feature, the mean value of all frames cannot be obtained. In the present invention, the mean value of the dynamic update is adopted to replace the mean value of the whole voice, and the formula is as follows:

$\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}}{\mathrm{<span\; class="notranslate">\; L</span>}}\underset{\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 14</span>}<span\; class="notranslate">\; )</span>$

Among them, L represents the number of filter channels, and λ _μ represents the forgetting coefficient, and the value range is (0～1). Using this mean value to normalize the short-term power spectrum of pure speech in each channel can eliminate the influence of multiplicative noise. The formula for this step is as follows:

$\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}\frac{\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ]</span>}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ]</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 15</span>}<span\; class="notranslate">\; )</span>$

Where k is an arbitrary constant, using such online processing, the online feature can achieve the effect of offline.

9. Perform equal loudness emphasis on the short-term power spectrum of pure speech after energy normalization processing

Their sound pressure intensity is different when different frequencies are equal. In order to compensate for this deviation of the human ear to frequency, equal loudness pre-emphasis is required on the power spectrum. Usually the center frequency of each channel is used as the frequency of this channel to compensate this channel, and there are various compensation formulas. The equal loudness weight formula adopted in the present invention is as follows:

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1.44</span>}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; 10</span>}}^{\mathrm{<span\; class="notranslate">\; 6</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}^{\mathrm{<span\; class="notranslate">\; 4</span>}}}{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1.6</span>}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; 10</span>}}^{\mathrm{<span\; class="notranslate">\; 5</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 9.61</span>}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; 10</span>}}^{\mathrm{<span\; class="notranslate">\; 6</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 16</span>}<span\; class="notranslate">\; )</span>$

Where w represents the frequency, and l represents the index of the channel, w _l is the frequency of the lth channel, that is, the center frequency of the lth channel.

The following formula is used for equal loudness emphasis on the short-term power spectrum of pure speech after energy normalization processing:

O[m,l]=U[m,l]·E(w _l ) (17)

where m, l are the indices of frame and channel respectively.

10. Exponential operation on the short-term power spectrum of pure speech after equal loudness emphasis

In order to better fit the human auditory model and convert the intensity into loudness, the power spectrum needs to be compressed nonlinearly. In the traditional PLP feature, the cubic root nonlinearity is used; in the traditional MFCC, the in a logarithmic non-linear fashion. In the feature extraction method of the present invention, an exponential nonlinear method is adopted. The formula is as follows:

L[m,l]=O[m,l] ^θ (18)

where θ is an exponential nonlinear parameter.

11. Inverse Fourier transform of the pure speech short-term power spectrum after exponential nonlinear transformation

The purpose of performing inverse Fourier transform on the pure speech short-time power spectrum after exponential nonlinear transformation is to obtain the cepstral coefficient of the speech signal, and then obtain the speech features. The inverse Fourier transform here uses the basic inverse Fourier transform method.

12. Find the cepstrum coefficient of the signal

In order to obtain the cepstral coefficients, in the method of the present invention, the Durbin recursive algorithm is first used to obtain the linear prediction coefficients, and then the corresponding cepstral coefficients are obtained according to the recursive formula by using the obtained linear prediction coefficients. The recursion formula is as follows:

${\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; =</span>\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; kc</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}\\ {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; p</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; kc</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ifn</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 19</span>}<span\; class="notranslate">\; )</span>$

Where a is the linear prediction coefficient, k is the reflection coefficient, they are all obtained by the Durbin recursive algorithm according to the autocorrelation equation of the Fourier inverse transform described in step 11, and n is the index to the spectral coefficient, p is the model Order.

13. Normalize the mean value of the cepstral coefficient

Although the energy normalization is performed in step 8, the mean normalization is still necessary, at least the mean normalization will not bring negative effects. Mean normalization is to calculate the average value of all frames for all dimensions of the cepstral coefficient, and then subtract the mean value of the spectral coefficient of the corresponding dimension from each dimension of the cepstral coefficient of each frame. Since the feature extraction method of the present invention is online, the mean is also the average of all frames before the current frame.

In the following, an example of a speech feature extraction method for robust speech recognition according to the present invention will be described in conjunction with the accompanying drawings. For speech with a sampling frequency of 16KHz, the specific introduction is as follows.

1. Pre-emphasize the speech signal, and the emphasis coefficient α is 0.97. The system function is shown in formula (1).

2. The voice frame length is 25ms, the frame shift is 10ms, add a Hamming window, and fill the tail of 400 points of a frame of voice with zeros to 512 points, and then use the fast Fourier transform to obtain the voice spectrum.

3. Utilize the obtained speech frequency spectrum to obtain the speech power spectrum according to the formula (2).

4. Use the Gamma-tone filter bank to process the power spectrum, the number of channels is 40, and the formula used is shown in (3).

5. Adopt the mode of frame average to ask for the power spectrum of medium time length, calculation formula is as above-mentioned formula (4), wherein M=2, promptly utilize the average power of current frame and its preceding two frames and following two frames as noise-containing speech The medium-duration power is used to replace the original medium-duration power of a single frame, and the duration is [(2M+1)-1]*10ms+25ms=65ms.

6. Perform asymmetric filter processing on the power spectrum to track the noise in the voice, and at the same time mask the power spectrum to obtain a pure voice power spectrum. In this step, the asymmetric low-pass filter anti-noise module including masking is calculated according to the formula in the implementation method. The formulas used are (5, 6, 7, 8, 9, 10, 11), and the values of the specific parameters are described as follows:

λ _a =0.999, λ _b =0.5

c=2

λ _t = 0.85, μ _t = 0.2

7. Carry out channel average to the ratio of pure speech and noisy speech power spectrum, to carry out smoothing, the formula that adopts is (12), and wherein N=4 (the channel number that promptly waits and sees forward and backward is 4 respectively), is about to The values of 9 channels are smoothed. Multiply the power spectrum ratio of the smoothed pure speech to the noisy speech with the short-term power spectrum of the noisy speech output by the filter bank, that is, formula (13), to obtain the short-term power spectrum of the pure speech.

8. Perform energy normalization processing on the short-term power spectrum of pure speech to eliminate multiplicative noise, as shown in formula (15). The mean value is dynamically estimated, as shown in formula (14), and the initial value is obtained from the statistics of the data set.

9. The equal loudness of the power spectrum is emphasized to make it conform to the auditory effect of the human ear.

10. Perform an exponential operation on the power spectrum to make it conform to human physiological characteristics, where the exponential nonlinear parameter θ is selected as 1/15.

11. The inverse Fourier transform of the power spectrum can be calculated using the basic formula of the inverse Fourier transform, because the number of points itself is small and the amount of calculation is not large.

12. When calculating the cepstral coefficient of the signal, the selected linear prediction coefficient is 12, and the cepstral coefficient is also 12. The formula is shown in (17).

13. Carry out mean value normalization processing to cepstral coefficient, finally obtain the speech feature of the method of the present invention.

The feature extraction method proposed by the present invention is compared with the effects of commonly used feature extraction methods:

Utilize feature extraction method described in the present invention to extract the feature of voice on 863 desktop voice collections, use the advanced anti-noise front end (AFE) feature extraction of PLP feature extraction method and European Telecommunications Standards Institute (ETSI) simultaneously the voice feature of 863 desktop voice collections . Using these three feature sets, the acoustic model is trained using HTK tools under the same conditions. Then, select 1000 pure reading recordings, add simulated white noise, and then use the above three feature extraction methods to extract features respectively. In addition, a set of random conversation recordings is marked, and 7072 clean recordings and 360 noise recordings are obtained, and the above three feature extraction methods are still used to extract speech features respectively.

The above acoustic model and its corresponding features are used for speech recognition, the language model uses the same 3-gram language model, and the recognizer uses the decoder in the HTK tool. Here, word error rate (WER) is used to evaluate speech recognition performance, wherein PNPLP is the name of the feature extraction algorithm of the present invention. The calculation formula of WER is as follows:

Under the test conditions of simulated white noise, the performance of various features is shown in Table 1. It can be seen from Table 1 that in the case of pure speech without noise, the PLP feature shows good performance, but as the noise increases, the PLP performance gradually deteriorates. The anti-noise feature (AFE) of the European Telecommunications Standards Institute can show certain effects on noise, but the anti-noise performance of the feature extraction method of the present invention is far better than the anti-noise algorithm of the European Telecommunications Standards Institute.

Table 2 is the experimental results of various feature extraction algorithms on the real test set. It can be seen from the attached table that the feature extraction method of the present invention has outstanding anti-noise performance, which is much better than the anti-noise performance of the European Telecommunications Standards Institute. In addition, the anti-noise feature extraction algorithm of the present invention has a slight decline compared with the classic PLP algorithm on the pure speech set, but compared with the anti-noise algorithm of the European Telecommunications Standards Institute, the anti-noise feature extraction algorithm of the present invention is still better a lot of.

Table 1

Table 2

WER PLP AFE PNPLP clear 11.64% 13.68% 12.07% noise 35.89% 34.21% 33.36%

The specific embodiments described above have further described the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (9)

1. for a Speech Feature Extraction for robust speech identification, it is characterized in that, the method comprises:

Step 1, obtain the power spectrum of voice signal;

Step 2, by obtained power spectrum by bank of filters processing, obtain the short-time rating spectrum of noisy speech;

Step 3, according to the short-time rating spectrum of obtained noisy speech, adopt the average mode of frame to ask for the medium duration power spectrum of noisy speech;

Step 4, the medium duration power spectrum of obtained noisy speech is carried out asymmetric filtering and shelters anti-noise, to obtain the medium duration power spectrum of clean speech;

Step 5, obtain the short-time rating spectrum of clean speech according to the short-time rating spectrum of the medium duration power spectrum of the medium duration power spectrum of described clean speech, noisy speech and noisy speech;

Step 6, the short-time rating spectrum of clean speech is carried out to energy normalized processing, to eliminate the property taken advantage of noise;

Step 7, offset and the loudness such as carry out except the short-time rating spectrum of the clean speech of the property taken advantage of noise and increase the weight of;

The short-time rating spectrum of the clean speech after step 8, reciprocity loudness increase the weight of is carried out index nonlinear operation;

Step 9, the short-time rating spectrum of having carried out the clean speech after index nonlinear operation is carried out to inverse fourier transform, to ask for cepstrum coefficient, cepstrum coefficient is carried out to average normalized, finally obtain phonetic feature;

Wherein, the frequency spectrum that obtains voice signal described in step 1 further comprises following content:

Step 11, to voice signal adopt formula (1) carry out pre-emphasis:

$y_t=\left\{\begin{array}{cc}{x}_t-\mathrm{\&alpha;}\&CenterDot;x_{t-1}& \mathrm{ift}>0\\ (1-\mathrm{\&alpha;})\&CenterDot;x_t& \mathrm{ift}=0\end{array}\right.---\left(1\right)$

Wherein α is called as pre emphasis factor, and x is speech sample point, and y is the speech sample point value after pre-emphasis, the index that t is sampled point;

Step 12, each frame of the voice after pre-emphasis is carried out to windowing, adopt Hanning window or hamming window, voice after windowing are carried out to short time discrete Fourier transform, comprise and adopt original Fourier to change formula, or utilize fast fourier transform to solve to the integer power of the voice zero padding to 2 after windowing;

Step 13, the real part after Short Time Fourier Transform and imaginary part are asked for respectively square and summation, to obtain the power spectrum of voice signal, as shown in formula (2):

P(w)＝Re[S(w)] ²+Im[S(w)] ² (2)

The power spectrum that wherein P (w) is voice signal, S (w) represents fourier spectra in short-term, Re[S (w)] and Im[S (w)] real part and the imaginary part of fourier spectra in short-term represented respectively.

2. the Speech Feature Extraction for robust speech identification according to claim 1, it is characterized in that, the bank of filters adopting in described step 2 is Mel bank of filters Mel-filter bank or Gamma-tone bank of filters, port number is chosen different numbers according to different wave filters, wherein, the short-time rating that utilizes the summation of Gamma-tone bank of filters to obtain noisy speech is composed, as shown in formula (3):

$P[m,l]=\underset{k=0}{\overset{(K/2)-1}{\mathrm{\&Sigma;}}}{\left|X\right[m,e^{{\mathrm{jw}}_k}\left]H_l\left(e^{{\mathrm{jw}}_k}\right)\right|}^2---\left(3\right)$

Wherein p[m, l] be the short-time rating spectrum of noisy speech, m and l represent respectively the index of frame and bank of filters passage, what K was Fourier transform counts, w _k=2 π/F _s, F _srepresent the sample frequency of voice signal, represent voice signal m frame the the amplitude of frequency, represent l passage the Gamma-tone filter value of frequency.

3. the Speech Feature Extraction for robust speech identification according to claim 1, is characterized in that, adopts the average mode of frame to ask for the medium duration power spectrum of noisy speech in step 3, as shown in formula (4):

$Q[m,l]=\frac{1}{2M+1}\underset{m^{\&prime;}=m-M}{\overset{m+M}{\mathrm{\&Sigma;}}}P[m^{\&prime;},l]---\left(4\right)$

Wherein Q[m, l] be the medium duration power spectrum of noisy speech, m and l represent respectively the index of frame and bank of filters passage, when M represents to ask for medium duration, the frame number sum got forward and backward respectively, P[m ', l] be the noisy speech short-time rating spectrum of the m ' frame.

4. the Speech Feature Extraction for robust speech identification according to claim 1, is characterized in that, described in step 4, the medium duration power spectrum of obtained noisy speech is carried out asymmetric filtering and shelter anti-noise specifically comprising following step:

Step 41, medium obtained noisy speech duration power spectrum is carried out to filtering through first asymmetric low-pass filter, and the Output rusults that medium described noisy speech duration power spectrum is deducted to described first asymmetric low-pass filter is to integrate; Wherein said first asymmetric low-pass filter as formula (5) represent:

$Q_{\mathrm{le}}[m,l]=\left\{\begin{array}{c}{\mathrm{\&lambda;}}_a{Q}_{\mathrm{le}}[m-1,l]+(1-{\mathrm{\&lambda;}}_a)Q[m,l],\mathrm{ifQ}[m,l]\&GreaterEqual;Q_{\mathrm{le}}[m-1,l]\\ {\mathrm{\&lambda;}}_b{Q}_{\mathrm{le}}[m-1,l]+(1-{\mathrm{\&lambda;}}_b)Q[m,l],\mathrm{ifQ}[m,l]<Q_{\mathrm{le}}[m-1,l]\end{array}\right.---\left(5\right)$

Wherein m and l represent respectively the index of frame and bank of filters passage, Q _le[m, l] is the output of described first asymmetric low-pass filter; Q[m, l] be the medium duration power spectrum of described noisy speech, λ _aand λ _bfor adjustable parameter, span is (0～1);

Step 42, the result after described integration is obtained to Q through a half-wave rectification block _o[m, l], by Q _o[m, l] sends into respectively masking block and second asymmetric low-pass filter processed, and described second asymmetric low-pass filter is identical with first asymmetric low-pass filter, and the output of second dissymetrical filter is as spectrum end power; The Q that wherein half-wave rectification block obtains _othe formula of [m, l] is expressed as follows:

Q _o[m，l]＝max(Q[m，l]-Q _le[m，l]，0) (6)

Step 43, the described Q obtaining through half-wave rectification block _o[m, l] obtains result Q after masking block is processed _tm[m, l], and Q _o[m, l] obtains result Q after second asymmetric low-pass filter processed _f[m, l], afterwards, described Q _tm[m, l] and described Q _f[m, l] is input to maximal value module and obtains result R _sp[m, l]; Wherein maximal value module is as shown in formula (7):

R _sp[m，l]＝max(Q _tm[m，l]，Q _f[m，l]) (7)

Step 44, determined the medium duration power spectrum R[m of clean speech by the first selective switch, l], described the first selective switch is as shown in formula (8):

$R[m,l]=\left\{\begin{array}{c}{R}_{\mathrm{sp}}[m,l],\mathrm{ifQ}[m,l]\&GreaterEqual;{\mathrm{cQ}}_{\mathrm{le}}[m,l]\\ Q_f[m,l],\mathrm{ifQ}[m,l]<{\mathrm{cQ}}_{\mathrm{le}}[m,l]\end{array}\right.---\left(8\right)$

Wherein c is adjustable parameter.

5. the Speech Feature Extraction for robust speech identification according to claim 4, is characterized in that, the operating process of described masking block comprises following content:

The described Q obtaining through half-wave rectification block _o[m, l] obtains Q through the MAX module of masking block _p[m, l], as shown in formula (9):

Q _p[m，l]＝max(λ _tQ _p[m-1，l]，Q _o[m，l]) (9)

Wherein λ _tbe Forgetting coefficient, span is (0～1), the output valve Q that masking block is last _tm[m, l] determined by the second selective switch, and described the second selective switch is as shown in formula (10):

$Q_{\mathrm{tm}}[m,l]=\left\{\begin{array}{cc}{Q}_o[m,l]& Q_o[m,l]\&GreaterEqual;{\mathrm{\&lambda;}}_t{Q}_p[m-1,l]\\ {\mathrm{\&mu;}}_t{Q}_p[m-1,l]& Q_o[m,l]<{\mathrm{\&lambda;}}_t{Q}_p[m-1,l]\end{array}\right.---\left(10\right)$

Wherein μ _tfor corresponding parameter, span is (0～1), the output valve Q of masking block _tmafter the selective switch that [m, l] describes through formula (8), final, the medium duration power spectrum of noisy speech is carried out to asymmetric filtering and the result of sheltering after anti-noise is R[m, l].

6. the Speech Feature Extraction for robust speech identification according to claim 1, is characterized in that, the short-time rating spectrum of obtaining clean speech in described step 5 comprises following content:

Step 51, calculates the ratio H[m of the medium duration power spectrum of obtained clean speech and the medium duration power spectrum of noisy speech, l], as shown in formula (11):

$H[m,l]=\frac{R[m,l]}{Q[m,l]}---\left(11\right);$

Wherein said R[m, l] be the medium duration power spectrum of clean speech, Q[m, l] be the medium duration power spectrum of noisy speech;

Step 52, carries out passage average, to obtain passage average weight H _s[m, l], as shown in formula (12):

$H_s[m,l]=\left(\frac{1}{l_2-l_1+1}\underset{i^{\&prime;}=l_1}{\overset{l_2}{\mathrm{\&Sigma;}}}H\right[m,l\left]\right)---\left(12\right)$

Wherein l ₂=min (l+N, L), l ₁=max (l-N, 1), L represents the number of filter channel, N represents to ask for passage average time, the total number of channels of looking around forward and backward;

Step 53, utilizes described passage average weight H _sthe short-time rating spectrum P[m of [m, l] modulation noisy speech, l], and obtain the short-time rating spectrum T[m of clean speech, l], as shown in formula (13):

T[m，l]＝P[m，l]H _s[m，l] (13)。

7. the Speech Feature Extraction for robust speech identification according to claim 1, is characterized in that, the short-time rating spectrum to clean speech in described step 6 is carried out energy normalized processing, as shown in formula (15):

$U[m,l]=k\frac{T[m,l]}{\mathrm{\&mu;}\left[m\right]}---\left(15\right)$

Wherein k is arbitrary constant, T[m, l] be the short-time rating spectrum of clean speech, μ [m] is as shown in formula (14):

$\mathrm{\&mu;}\left[m\right]={\mathrm{\&lambda;}}_{\mathrm{\&mu;}}\mathrm{\&mu;}[m-1]+\frac{1-{\mathrm{\&lambda;}}_{\mathrm{\&mu;}}}{L}\underset{l=0}{\overset{L-1}{\mathrm{\&Sigma;}}}T[m,l]---\left(14\right)$

Wherein L represents the number of filter channel, λ _μrepresent Forgetting coefficient, span is (0～1).

8. the Speech Feature Extraction for robust speech identification according to claim 1, is characterized in that, in described step 7, the described spectrum of the short-time rating to clean speech the loudness such as is carried out and increased the weight of as shown in formula (17):

O[m，l]＝U[m，l]·E(w _l) (17)

Wherein m, l is respectively the index of frame and passage, U[m, l] for voice are through anti-noise short-time rating spectrum after treatment, E (w _l) as shown in formula (16):

$E\left(w_l\right)=\frac{(w_l^2+1.44\&times;{10}^6)w_l^4}{{(w_l^2+1.6\&times;{10}^5)}^2\&times;(w_l^2+9.61\&times;{10}^6)}---\left(16\right)$

Wherein w represents frequency, w _lit is the frequency of l passage.

9. the Speech Feature Extraction for robust speech identification according to claim 1, it is characterized in that, in described step 8, the short-time rating spectrum of the clean speech after described reciprocity loudness increases the weight of is carried out index nonlinear operation as shown in formula (18):

L[m，l]＝O[m，l] ^θ (18)

Wherein θ is the nonlinear parameter of index, O[m, l] be the short-time rating spectrum of the clean speech of the loudness such as described after increasing the weight of.