CN104091593A - Speech endpoint detection algorithm adopting perceptual speech spectrum structure boundary parameters - Google Patents

Abstract

The invention belongs to the field of speech recognition and discloses a speech endpoint detection algorithm using perceptual spectrum structure boundary parameters (PSSB). After the speech enhancement based on the auditory perception characteristics of the noisy speech, according to the difference between the continuous distribution characteristics of the speech signal and the random distribution characteristics of the residual noise, two-dimensional enhancement is performed on the time-frequency spectrum of the enhanced speech, Thus further highlighting the spectral structure of continuous distribution of pure speech. Through the two-dimensional boundary detection of the enhanced speech spectrum structure, the PSSB parameters are proposed and used for endpoint detection. The experimental results show that the endpoint detection algorithm using PSSB parameters can detect the endpoint of the speech more effectively in various signal-to-noise ratio environments from -10dB to 10dB of white noise. Under the extremely low SNR of -10dB, the proposed method still has a correct rate of 75.2%.

Description

Speech Endpoint Detection Algorithm Using Perceptual Spectral Structure Boundary Parameters

technical field

The invention belongs to the field of speech recognition and relates to a speech endpoint detection algorithm, in particular to a speech endpoint detection algorithm using perceptual language spectrum structure boundary parameters. the

Background technique

As the basis of speech recognition and speaker recognition, correct and effective endpoint detection can greatly improve the recognition rate of speaker recognition system and speech recognition system. In the environment of high signal-to-noise ratio in the laboratory, the traditional endpoint detection algorithm can detect the voice endpoint very well. However, under low SNR environment, the performance of most endpoint detection algorithms drops sharply. the

In recent years, many scholars have conducted research on noise-robust endpoint detection. Ganapathiraju (A. Ganapathiraju, et al. Comparison of Energy-Based Endpoint Detectors for Speech Signal Processing . In Proc. lEEE Publications, 1996; 500-503) et al. adopted a method combining short-term energy and short-term zero-crossing rate ( Energy and Zero-Crossing Rate, EZCR) for endpoint detection research. Compared with traditional energy methods, this method has better robustness for endpoint detection. However, this method cannot work in an environment with a lower signal-to-noise ratio. Chen Zhenbiao et al. (Chen Zhenbiao, Xu Bo. Research on Optimal Speech Endpoint Detection Algorithm Based on Subband Energy Features. Acoustica Sinica, 2005;30(2):171-176) studied subband Amplitude [Sub-Band Amplitude, SBA] and energy, and use a detection algorithm that combines more distinguishable and anti-noise multiple sub-band energies and optimized edge detection commonly used in image processing to perform endpoint detection, making endpoint detection Performance in complex noise environments has been significantly improved. In addition, Zhang et al. ( Xueying Zhang , et al. A Speech Endpoint Detection Method Based on Wavelet Coefficient Variance and Sub-Band Amplitude Variance. . In Proc. lEEE ICICIC, 2006; 105-109) proposed a method using wavelet coefficients ( Wavelet Coefficient, WC) method uses wavelet analysis method for endpoint detection. Since this method can analyze signals at various scales, it can distinguish speech segments and noise segments to a certain extent. Wu et al. (Bing-Fei Wu, Kun-Ching Wang. Robust Endpoint Detection Algorithm Based on the Adaptive Band-Partitioning Spectral Entropy in Adverse Environments. IEEE Transactions on Speech and Audio Processing, 2005; 13(5):562-7 The adaptive band-partitioning spectral entropy (Adaptive Band-Partitioning Spectral, ABSE) method is used for endpoint detection. This method can distinguish the speech sub-band signal and noise very well, and achieves a good accuracy rate of endpoint detection in the environment containing noise. Li (Q.Li, et al. A Robust real-time endpoint detector with energy normalization for ASR in adverse environments. International Conference on Acoustics Speech and Signal Processing, 2001; 574-577) draws on the method of optimizing edge detection in image processing It is used for endpoint detection of speech, and a filter plus three-state decision logic is used for endpoint detection, so there is no need to adjust the threshold in the case of different signal-to-noise ratios. This method combines the algorithm of image processing, which plays a very good auxiliary role in the detection of endpoints. However, none of the above methods can obtain a high accuracy rate of endpoint detection in a low signal-to-noise ratio environment. the

Contents of the invention

Technical problem to be solved: In the environment of low signal-to-noise ratio, the correct rate of endpoint detection of conventional endpoint detection methods is very low. the

Technical solution: Aiming at the different characteristics of speech signals and noise signals in the two-dimensional space of time-frequency domain under low signal-to-noise ratio, combined with the speech enhancement algorithm based on auditory perception characteristics, the perceptual spectrum structure boundary parameter PSSB (Perception Spectrogram Structure Boundary) is proposed ), and use it for endpoint detection. Firstly, speech enhancement based on auditory masking characteristics is performed on speech with low signal-to-noise ratio. Compared with traditional speech enhancement algorithms, this method more effectively preserves the speech components perceivable by the human ear. On this basis, considering the continuous distribution characteristics of the pure speech spectrum on the time axis in the two-dimensional level, two-dimensional enhancement is performed on the noisy speech, so that the spectral structure of the speech is further highlighted, and the noise is suppressed at the same time. spectral structure. Finally, the two-dimensional boundary of the continuous distribution of the pure speech spectral structure is found, and the PSSB parameters are proposed for endpoint detection. the

1. Speech enhancement based on auditory perception characteristics

In a low SNR environment, most endpoint detection algorithms cannot detect voice endpoints well, or even fail completely. Humans, on the other hand, can recognize speech segments in noisy environments. In a noisy environment, the auditory perception characteristics of the human ear play an important role. Using the auditory masking characteristic in the auditory perception characteristics of the human ear can suppress noise to a certain extent and retain more speech components. The PSSB parameters proposed by the present invention first adopt the speech enhancement based on the auditory masking characteristic, and suppress the noise as much as possible on the basis of protecting the speech. In this speech enhancement method, the most important thing is to calculate the masking threshold. The calculation of the masking threshold and the speech enhancement system are as follows:

(1) Bark threshold power spectrum

The speech signal x(n) is converted into a frequency domain signal by fast Fourier transform (FFT) , the signal power spectrum is:

(1)

The Bark power spectrum is:

$B_i=\underset{k=b_{\mathrm{li}}}{\overset{b_{\mathrm{hi}}}{\mathrm{\Σ}}}P\left(k\right)---\left(2\right)$ in Indicates the energy of the i-th Bark band, Indicates the lowest frequency of segment i, Indicates the highest frequency of segment i.

(2) Diffused Bark domain power spectrum

Introduce the spread function , which is a matrix that satisfies the condition:

(3)

The definition formula is as follows:

(4)

Indicates the difference between the band numbers of the two bands.

${\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; j</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

(3) Offset function of masking energy and masking threshold calculation

(6)

$T_i={10}^{{\log }_{10}\left(C_i\right)-(o_i/10)}---\left(7\right)$ The value is between 0 and 1, determined by the speech content. is the masking threshold of the i-th Bark band, which is renamed as , where b has the same meaning as the previous i.

and the threshold of the quiet hearing threshold:

(8)

In comparison, the maximum value is taken as the masking threshold for the final fitting. in for The corresponding Bark masking curve.

(4) Adjustment of spectrum subtraction and subtraction parameters

The gain function used by the spectral subtraction algorithm is as follows:

$h\left(k\right)=\left\{\begin{array}{cc}{(1-\mathrm{\&alpha;}\&CenterDot;{\left[\frac{\left|\mathrm{D.}\left(k\right)\right|}{\left|Y\left(k\right)\right|}\right]}^{\mathrm{\&gamma;}})}^{1/\mathrm{\&gamma;}},& {\left[\frac{\left|\mathrm{D.}\left(k\right)\right|}{\left|Y\left(k\right)\right|}\right]}^{\mathrm{\&gamma;}}<\frac{1}{\mathrm{\&alpha;}+\mathrm{\&beta;}}\\ {(\mathrm{\&beta;}\&Center\; Dot;{\left[\frac{\left|\mathrm{D.}\left(k\right)\right|}{\left|Y\left(k\right)\right|}\right]}^{\mathrm{\&gamma;}})}^{1/\mathrm{\&gamma;}},& \mathrm{else}\end{array}\right.---\left(9\right)$ First calculate the noise masking threshold of different Bark domains of each frame of speech, and then get the adaptive subtraction parameter according to the noise masking threshold , : If the masking threshold is high, the residual noise will be masked naturally to make it inaudible to the human ear. In this case, the subtraction parameters take their minimum value; when the masking threshold is low, the residual noise will have a great influence on the human ear. large, it is necessary to reduce it. For each frame m, the masking threshold The minimum value of and the subtraction parameter per frame and about the maximum value of . The application of the subtraction parameter has the following relation:

,

(10)

in, and respectively minimum and maximum values of . , and , are the parameters , minimum and maximum values of . when hour, ;when hour, . In the formula and are the minimum and maximum masking thresholds obtained frame by frame, respectively. In the experiment, we set the values of each parameter as follows:

(5) Real-time noise power spectrum estimation

Speech enhancement requires noise spectrum estimation methods with particularly high real-time performance. Noise power spectrum estimation method based on constrained variance spectral smoothing and minimum value tracking is adopted. The core of the algorithm is the variance-constrained smoothing filter, which controls the variance of the short-term smooth power spectrum, making the tracking of the minimum more accurate. The noise spectrum estimated by this method can track the noise mutation in time without obvious noise spectrum delay, and its accuracy is better than that estimated by other methods.

(6) Speech Enhancement System

Get an adaptive subtraction parameter according to the masking threshold ,. The voice enhancement system is shown in Figure 1.

2 Two-dimensional enhancement of speech

After the speech with low signal-to-noise ratio is enhanced, the noise and the speech are attenuated at the same time due to the effect of spectral subtraction. However, because voiced segments in speech contain structures such as formants with high energy, in the two-dimensional time-frequency domain, the low-frequency region of the speech spectrum has a high signal-to-noise ratio even under noise interference. And these structures containing higher speech energy are usually distributed continuously in time. Therefore, as long as we find these continuously distributed high-energy regions in the two-dimensional speech spectrum of the speech signal, and thus find out the connected unvoiced segments, we can get the start and end endpoints of the speech. Boundary detection, in our approach, is an algorithm that finds continuously distributed 2D data structures.

However, no matter whether the speech signal with low SNR is speech enhanced or not, the noise (residual music noise after speech enhancement) will leave the boundary of the noise spectral structure in the boundary detection. The spectral structure of pure speech will be confused by the spectral structure of noise, which will have a great interference effect on finding the spectral structure of pure speech. As shown in Figure 2 and Figure 3. the

Figure 2 is a spectrogram of a speech containing -5dB white noise. As can be seen in the figure, the continuous black horizontal stripes are speech signals (in the high-frequency band, the speech signal with lower energy has been masked by noise, and the formant structure in the high-frequency region cannot be seen from the spectrogram), And the black snowflake-like background is white noise. Figure 3 is a spectrogram after speech enhancement. After speech enhancement, the noise is greatly weakened, but there are still residual musical noises of different strengths. The present invention divides these residual noises into residual noises with stronger energy and residual noises with weaker energy, as shown in FIG. 3 . These noises will greatly interfere with the endpoints for obtaining voice. Therefore, before obtaining the speech endpoint, the present invention performs two-dimensional enhancement on the speech, including a two-dimensional noise erosion algorithm and a two-dimensional speech expansion algorithm, for the difference between the spectral structure of the residual noise and the pure speech. the

2D Noise Erosion Algorithm

Among the enhancement processing algorithms for 2D data, corrosion algorithms can weaken or eliminate specific 2D structures. We found that in the speech spectrum after speech enhancement, residual noise with weak energy (dark snowflake-like structure) is usually randomly distributed, as shown in Figure 3. And they have smaller size and energy. Although these structures are not as strong as the white noise in Figure 3, they still interfere with the spectral structure boundary of pure speech. Aiming at the above characteristics, the present invention proposes a two-dimensional noise erosion algorithm for weakening such a two-dimensional structure.

The two-dimensional noise erosion algorithm for the speech spectrum is determined by the following process. First, perform a short-time Fourier transform on the speech, and the spectrum of each frame Calculated by the following formula:

(11)

is the speech signal of the mth frame, is the frequency spectrum of the speech signal of the mth frame. N is the length of the frame and the number of short-time Fourier transform points. It is a Hamming window. The speech signal power spectrum of each frame can be expressed as:

(12)

That is, it is defined as the spectrum of the speech signal.

right The 2D noise erosion of is defined as:

(13)

in is a structural element, yes domain of definition, yes domain of definition. translation parameter gotta be within the domain of definition, and gotta be within the domain of definition. The effect of two-dimensional noise erosion on the signal is two-fold: (1) if all elements are positive, the output signal tends to be weaker than the original signal; (2) in the input spectral signal, the noise spectral structure if Similar to structural elements, it will be attenuated, and the degree of attenuation depends on the shape of the spectral structure of the noise and the shape of the structural elements.

In the spectral structure of speech, erosion algorithms attenuate both noise and speech. The purpose of the two-dimensional noise erosion algorithm proposed by the present invention is to attenuate the noise relatively more and preserve the speech better. Structural elements of the two-dimensional noise erosion algorithm for the structural shape of the residual noise spectrum with weak energy is defined as:

(14)

Such structural elements Compare the spectral structure (smaller dots) closer to the less energetic residual noise. So with the structuring element Carrying out two-dimensional noise erosion on the speech spectrum can weaken this noise to a certain extent.

Two-dimensional Speech Expansion Algorithm

After the speech is passed through the two-dimensional noise erosion algorithm, the residual noise with weak energy is well suppressed. However, due to the similarity in energy between the residual noise with strong energy (as shown in Figure 3) and the pure speech, if the corrosion is excessive, the two-dimensional structure of the pure speech will be weakened at the same time. The expansion algorithm can enhance the two-dimensional spectral structure similar to the structural elements, and relatively weaken the dissimilar two-dimensional spectral structure. Therefore, the present invention proposes a two-dimensional speech expansion algorithm for the difference between residual noise with strong energy and pure speech structure. The present invention defines a structural element as a structure similar to a continuous distribution of pure speech. In this way, the noise structure can be relatively suppressed.

Results for 2D noise erosion , two-dimensional speech expansion algorithm is defined by:

(15)

in is a structural element, yes domain of definition, yes domain of definition. Theoretically, it can be considered that the structure element is translated at all positions in the spectrum, the value of the structure element is added to the value of the two-dimensional signal, and the maximum value is calculated. Two-dimensional speech dilation on a speech signal has a dual effect: (1) if all elements are positive, the output signal tends to be stronger than the original signal; (2) whether a certain structure in the input spectral signal is Relative enhancement, depending on the value and shape of the structuring elements used for dilation.

The expansion algorithm, while enhancing the speech structure, will also enhance the corresponding noise structure. The purpose of the two-dimensional speech expansion algorithm proposed by the present invention is to enhance the speech structure as much as possible and relatively suppress the noise structure. The spectral structure of voiced sounds of pure speech signals is usually a long strip extending along the time axis, while the spectral structure of residual noise with strong energy is usually square or circular in different sizes, as shown in Figure 3. Therefore, the structure element is defined as a strip shape extending along the time axis, so as to enhance all similar structures, and at the same time, it can relatively weaken noise structures with different structures. the

Therefore, the structural elements in the two-dimensional speech expansion algorithm is defined as the following shape:

(16)

here It is a horizontal structural element that extends along the time direction. All structures similar to it will be enhanced. Since the spectral structure of pure speech is usually distributed continuously in time, it is similar to , so the structure of pure speech is strengthened. The spectral structure of residual noise with strong energy is usually in the shape of large round or square dots, and its structure is relatively weakened.

3 Perceptual Spectral Structure Boundary (PSSB) Parameters and Endpoint Detection Algorithm

3.1 Perceptual Spectral Structure Boundary (PSSB) Parameters

The present invention considers the continuous distribution characteristics of the pure speech spectrum on the time axis on the two-dimensional level, and performs two-dimensional enhancement on the noisy speech, so that the spectral structure of the speech is further highlighted, and the spectral structure of the noise is suppressed at the same time . Afterwards, the present invention will find out the spectral structure boundary of the continuous distribution of pure speech, and propose a perceptual spectral structure boundary parameter PSSB for endpoint detection.

For the perceptual spectral structure boundary parameter PSSB, the boundary information of the spectral structure must be solved first. Boundary detection is an important method to solve the boundary of 2D structures. The boundary of a continuous two-dimensional signal can be represented by a gradient determined by the first derivative. The present invention uses the neighborhood model in the formula (17) to approach the result of speech two-dimensional enhancement gradient.

                                          (17)

is the center point of this neighborhood model. The gradient of the central neighborhood can be expressed by the following formula:

(18)

and Determined by formula (19) and formula (20):

                                              (19)

    (20)

that is The boundary, which can describe the boundary information of the continuous distribution of the speech signal in the noisy speech spectrum.

by right And the analysis of the speech spectrum, we found that in the environment of low SNR, the signal and spectral features of the high-frequency region of the speech are masked by the noise, while in the low-frequency region, the spectral structure of the voiced segment of the speech is still relative to the noise. Very high energy, with solvable spectral boundaries. And the lower the frequency, the more obvious this phenomenon is. This is because the energy of the speech voiced segment is mainly concentrated in the first few formants of the middle and low frequencies. Therefore, after obtaining the boundary of the speech spectrum After that, on the frequency axis of each frame of the spectrum for all The weighted summation is performed to make the low-frequency region get a higher weight, so as to obtain the perceptual spectral structure boundary parameter PSSB.

The perceptual spectral structure boundary parameter PSSB is proposed as follows:

(twenty one)

in is the PSSB parameter of the mth frame, and M is the total number of frames.

PSSB parameters It can well reflect the relative content of the speech voiced segment signal in a frame, and has good robustness to noise.

3.2 Voice endpoint detection

Voiced segments in speech usually have a longer continuous distribution time. The unvoiced segment has two distribution types: (1) the unvoiced sound is distributed in the middle of the speech segment; (2) the unvoiced sound is distributed at the beginning of the speech segment.

Through experiments, it is found that the unvoiced sound in the middle of the speech segment can be well recognized as a speech segment (PSSB parameter is greater than the threshold 0.5). This is because the unvoiced sound in the middle of a phonetic word is usually shorter, and what the present invention adopts is the frame shifting method of overlapping 50%. This method can combine the unvoiced sound in the middle of the word and the voiced sound next to it for spectral analysis, so that the information of the voiced sound frame next to it can be reflected in the unvoiced sound frame. the

However, as the signal-to-noise ratio decreases, especially below 0dB, the PSSB discrimination characteristic of unvoiced sounds at the beginning of the speech segment weakens (small value). If only a fixed threshold is used to divide the endpoints, the performance of unvoiced sound detection will drop sharply. However, although the PSSB of unvoiced sounds is relatively small relative to voiced sounds, it usually still has some PSSB distinguishing properties (small but not zero). Therefore, the present invention adopts a detection method aimed at the distribution characteristics of speech continuity, so as to treat voiced segments and unvoiced segments at endpoints differently. The specific endpoint detection method is as follows:

(1) First detect the speech segment whose PSSB parameter is greater than the threshold a and is continuously distributed in m frames, and this segment is the detected voiced segment.

(2) Based on this segment, all segments connected with this segment and continuously greater than or equal to the threshold b are defined as speech segments. The value of the threshold b is small. In the experiment, the value of b is between 0.01 and 0.05, which has better recognition results. In this way, unvoiced segments with smaller PSSB values can be identified. the

(3) The start and end points of this speech segment are the speech endpoints. the

After experimental testing, for white noise, when a=0.5, b=0.01, m=20, the performance of the system is better. the

The block diagram of the endpoint detection algorithm of the present invention is shown in FIG. 4 . the

Beneficial effect:

The experiments were designed under different signal-to-noise ratio environments. The input speech with low signal-to-noise ratio is 16k samples and 16-bit quantization. Using the Hamming window, the frame length is 256, and the frame shift is 128. The speech is selected from the TIMIT speech database, and the white noise is from the NoiseX-92 noise database. Figure 5 is a waveform diagram of a speech example (artists) in the database, and Figure 6 is a speech waveform with a low SNR of -10dB by adding white noise.

In Fig. 5, the starting point of the voice is the 40th frame, and the ending point is the 87th frame. And when the voice signal is added with white noise to make the signal-to-noise ratio reach -10dB, the voice signal has been completely submerged in the white noise. Traditional endpoint detection algorithms cannot effectively extract speech endpoints from such speech signals. the

Figure 7 is the spectrogram of pure speech examples (artists), Figure 8 is the spectrogram of this low SNR speech, and Figure 9 is the spectrogram after speech enhancement based on auditory masking characteristics. the

It can be seen from Figure 8 that most of the spectral structure of the speech at -10dB low SNR has been submerged by the noise, and only the formant structure in the low frequency region can be distinguished from the noise. After speech enhancement, it can be seen from Figure 9 that the noise signal and the speech signal are weakened by speech enhancement at the same time, and there are still randomly distributed music noises. This is due to the inherent characteristics of the spectral subtraction algorithm itself. the

If the boundary of the spectrum is obtained directly from the spectrum in Figure 9, it is still difficult to distinguish noise from speech. Therefore, it is necessary to do two-dimensional enhancement in the speech spectrum. As shown in Figure 10 and Figure 11. the

Figure 10 is the result of Figure 9 after the two-dimensional noise erosion algorithm. It can be seen from FIG. 9 that, except for the residual noise with strong energy and the formant structure of speech at low frequencies, other residual noises are suppressed to a certain extent. Fig. 11 is the result of performing the two-dimensional speech expansion algorithm on the spectral structure of the speech in Fig. 10 . It can be seen that the noise spectral structure with strong energy of random distribution is relatively weakened. The spectral structure of speech is relatively enhanced. the

Afterwards, the boundary detection of Figure 11 is performed, as shown in Figure 12. It can be seen that between frame 40 and frame 85, the speech spectrum boundary structure in the low-frequency region is well solved. However, in non-speech regions, a lot of boundary structures with mid- and high-frequency noise are represented due to the 2D structure with a small amount of noise still remaining. This is not expected to be seen. Therefore, in the PSSB parameters, the boundary structure in the low-frequency region is given higher weight. In this way, speech and noise are well distinguished. Figure 13.

FIG. 13 is the PSSB parameters obtained from FIG. 12 . Obviously, in the case of -10dB, the PSSB parameters of the speech signal can still have outstanding distinguishing characteristics on the time axis. When doing endpoint detection, the PSSB parameter is tested for continuity. If the value of the PSSB parameter is continuously greater than 0, and the number of frames continuously greater than the threshold 0.5 is greater than 20 frames, then the PSSB parameter whose value is continuously greater than 0 is judged as voice part.

In the experiment, the endpoint detection algorithm ( PSSB ) of the present invention is compared with other four endpoint detection algorithms, and their accuracy rates are compared. The four methods are: 1. Energy-short time zero crossing rate (EZCR); 2. Subband amplitude method (SBA); 3. Wavelet coefficient method (WC); 4. Subband spectral entropy method (ABSE). The present invention selects 70 words in the TIMIT speech database as the object of endpoint detection, and performs endpoint detection 3 times for each word. Add the white noise in the NoiseX-92 noise database according to a certain weight to get speech with different SNR. We set an endpoint detection with an error of less than 4 frames as the correct result. Define endpoint detection accuracy rate = correct result/total number of speech segments used for endpoint detection. Table 1 and Figure 14 show the accuracy of endpoint detection for various algorithms at different SNRs.

the

Table 1 The correct rate of endpoint detection under different signal-to-noise ratios (%)

"*" in Table 1 indicates that the algorithm fails under this condition, and we believe that the correct rate is zero. It can be seen from Table 1 and Figure 14 that in the case of 10dB, the correct rate of endpoint detection of the three traditional methods of EZCR, SBA and WC is already lower than 86%. When the SNR is below zero, the three methods fail completely, indicating that these methods are not robust to noise. The accuracy rate of the ABSE method is relatively high, because this method also analyzes the high-energy components of pure speech and makes endpoint detection. Compared with ABSE, the method using PSSB parameters in the present invention has a higher endpoint recognition rate. In the case of -10dB, there is still a correct recognition rate of 75.2%.

the

Description of drawings:

Fig. 1 is the speech enhancement system based on auditory characteristics;

Fig. 2 contains the spectrogram of -5dB white noise speech;

Spectrogram after Fig. 3 speech enhancement;

Fig. 4 is the endpoint detection algorithm adopting PSSB parameter;

Fig. 5 is pure voice;

Fig. 6 is -10dB low SNR voice;

Fig. 7 is pure speech signal spectrogram;

Fig. 8 is -10dB low SNR speech signal spectrogram;

Fig. 9 is speech enhancement result;

Fig. 10 is the spectrogram after two-dimensional noise erosion algorithm;

Fig. 11 is the spectrogram after two-dimensional speech expansion algorithm;

Fig. 12 is spectrum boundary;

Figure 13 shows the PSSB parameters and endpoint detection

Figure 14 is a comparison of endpoint detection results.

the

Detailed ways

Example 1

The first step: speech enhancement based on auditory perception characteristics; adopt speech enhancement based on auditory masking characteristics, suppress noise as much as possible on the basis of protecting speech; the calculation of masking threshold and the speech enhancement system in the described speech enhancement method are as follows:

ⅰ. Bark threshold power spectrum

The speech signal x(n) is converted into a frequency domain signal by fast Fourier transform (FFT) , the signal power spectrum is:

(1)

The Bark power spectrum is:

$B_i=\underset{k=b_{\mathrm{li}}}{\overset{b_{\mathrm{hi}}}{\mathrm{\&Sigma;}}}P\left(k\right)---\left(2\right)$ in Indicates the energy of the i-th Bark band, Indicates the lowest frequency of segment i, Indicates the highest frequency of the i segment;

ⅱ. Diffused Bark Domain Power Spectrum

Introduce the spread function , which is a matrix that satisfies the condition:

(3)

The definition formula is as follows:

(4)

Indicates the difference between the band numbers of the two bands;

$C_i=\underset{j=1}{\overset{j_{\max }}{\mathrm{\&Sigma;}}}{S}_{\mathrm{ij}}\&CenterDot;B_i,i=\mathrm{1,2}...i_{\max }---\left(5\right)$ ⅲ. Offset function of masking energy and masking threshold calculation

(6)

$T_i={10}^{{\log }_{10}\left(C_i\right)-(o_i/10)}---\left(7\right)$ The value is between 0 and 1, depending on the voice content; is the masking threshold of the i-th Bark band, which is renamed as , where the meaning of b is the same as the previous i;

and thresholds for quiet hearing:

(8)

Compare and take the maximum , as the masking threshold for the final fit; where for The corresponding Bark masking curve;

ⅳ. Spectral subtraction and adjustment of subtraction parameters

The gain function used by the spectral subtraction algorithm is as follows:

$h\left(k\right)=\left\{\begin{array}{cc}{(1-\mathrm{\&alpha;}\&CenterDot;{\left[\frac{\left|\mathrm{D.}\left(k\right)\right|}{\left|Y\left(k\right)\right|}\right]}^{\mathrm{\&gamma;}})}^{1/\mathrm{\&gamma;}},& {\left[\frac{\left|\mathrm{D.}\left(k\right)\right|}{\left|Y\left(k\right)\right|}\right]}^{\mathrm{\&gamma;}}<\frac{1}{\mathrm{\&alpha;}+\mathrm{\&beta;}}\\ {(\mathrm{\&beta;}\&Center\; Dot;{\left[\frac{\left|\mathrm{D.}\left(k\right)\right|}{\left|Y\left(k\right)\right|}\right]}^{\mathrm{\&gamma;}})}^{1/\mathrm{\&gamma;}},& \mathrm{else}\end{array}\right.---\left(9\right)$ First calculate the noise masking threshold of different Bark domains of each frame of speech, and then get the adaptive subtraction parameter according to the noise masking threshold , : If the masking threshold is high, the residual noise will be masked naturally to make it inaudible to the human ear. In this case, the subtraction parameters take their minimum value; when the masking threshold is low, the residual noise will have a great influence on the human ear. large, it is necessary to reduce it; for each frame m, the masking threshold The minimum value of and the subtraction parameter per frame and is related to the maximum value; the application of the subtraction parameter has the following relationship:

,

(10)

in, and respectively The minimum and maximum values of ; , and , are the parameters , The minimum and maximum values of ; when hour, ;when hour, ; where and They are the minimum and maximum values of the masking threshold obtained frame by frame; in the experiment, we set the values of each parameter as follows:

v. Real-time noise power spectrum estimation; using a noise power spectrum estimation method based on constrained variance spectral smoothing and minimum value tracking.

ⅵ. Speech enhancement system; according to the masking threshold, the adaptive subtraction parameter, ;

The second step: two-dimensional enhancement of speech;

2.1 Two-dimensional noise erosion algorithm

The two-dimensional noise erosion algorithm for the speech spectrum is determined by the following process; first, the short-time Fourier transform is performed on the speech, and the spectrum of each frame Calculated by the following formula:

(11)

is the speech signal of the mth frame, is the frequency spectrum of the m- th frame speech signal; N is the length of the frame and the number of short-time Fourier transform points; Is the Hamming window; the speech signal power spectrum of each frame can be expressed as:

(12)

That is, it is defined as the spectrum of the speech signal;

right The 2D noise erosion of is defined as:

(13)

in is a structural element, yes domain of definition, yes domain of definition; translation parameters gotta be within the domain of definition, and gotta be within the domain of definition;

Structural elements of the two-dimensional noise erosion algorithm for the structural shape of the residual noise spectrum with weak energy is defined as:

(14)

2.2 Two-dimensional Speech Expansion Algorithm

Results for 2D noise erosion , two-dimensional speech expansion algorithm is defined by:

(15)

in is a structural element, yes domain of definition, yes domain of definition;

Therefore, the structural elements in the two-dimensional speech expansion algorithm is defined as the following shape:

(16)

Step 3: Perceptual Spectral Structure Boundary (PSSB) Parameters and Endpoint Detection Algorithm

3.1 Perceptual Spectral Structure Boundary (PSSB) Parameters

The present invention uses the neighborhood model in the formula (17) to approach the result of speech two-dimensional enhancement gradient;

                                          (17)

is the center point of this neighborhood model; and the gradient of the center neighborhood can be expressed by the following formula:

(18)

and Determined by formula (19) and formula (20):

                                              (19)

    (20)

that is The boundary, which can describe the boundary information of the continuous distribution of the speech signal in the noisy speech spectrum.

The perceptual spectral structure boundary parameter PSSB is proposed as follows:

(twenty one)

in is the PSSB parameter of the mth frame, and M is the total number of frames;

3.2 Voice endpoint detection

A detection method aimed at the distribution characteristics of speech continuity is used to distinguish between voiced segments and unvoiced segments at the endpoints; the specific endpoint detection method is as follows:

(1) at first detect the speech segment that PSSB parameter is greater than threshold value a and distributes m frame continuously, and this segment is the voiced sound segment that detects;

(2) Based on this segment, all segments that are connected with this segment and are continuously greater than or equal to the threshold b are defined as speech segments; the value of the threshold b is smaller, and in the experiment, the value of b is between 0.01 and 0.05. have better recognition results. In this way, unvoiced segments with smaller PSSB values can be identified;

(3) The start and end points of this speech segment are the speech endpoints.

The experimental design is in different SNR environments; the input speech with low SNR is 16k samples, 16-bit quantization; using the Hamming window, the frame length is 256, and the frame shift is 128; the speech is selected from the TIMIT speech database, and the white noise is from NoiseX- 92 Noise Database. the

Claims (5)

1. A speech endpoint detection algorithm using perceptual spectrum structure boundary parameters, characterized in that the algorithm steps are as follows: (1) speech enhancement based on auditory perception characteristics; (2) two-dimensional enhancement of speech, including two-dimensional Noise erosion algorithm and two-dimensional speech expansion algorithm; (3) Perceptual Spectral Structure Boundary (PSSB) parameters and speech endpoint detection. 2. a kind of speech endpoint detection algorithm that adopts perceptual spectrum structure boundary parameter according to claim 1, it is characterized in that described described algorithm step is as follows: The first step: speech enhancement based on auditory perception characteristics; adopt speech enhancement based on auditory masking characteristics, suppress noise as much as possible on the basis of protecting speech; the calculation of masking threshold and the speech enhancement system in the described speech enhancement method are as follows: ⅰ. Bark threshold power spectrum The speech signal x(n) is converted into a frequency domain signal by fast Fourier transform (FFT) , the signal power spectrum is: (1) The Bark power spectrum is: in Indicates the energy of the i-th Bark band, Indicates the lowest frequency of segment i, Indicates the highest frequency of the i segment; ⅱ. Diffused Bark Domain Power Spectrum Introduce the spread function , which is a matrix that satisfies the condition: (3) The definition formula is as follows: (4) Indicates the difference between the band numbers of the two bands; ⅲ. Offset function of masking energy and masking threshold calculation (6) The value is between 0 and 1, depending on the voice content; is the masking threshold of the i-th Bark band, which is renamed as , where the meaning of b is the same as the previous i; and thresholds for quiet hearing: (8) Compare and take the maximum , as the masking threshold for the final fit; where for The corresponding Bark masking curve; ⅳ. Spectral subtraction and adjustment of subtraction parameters The gain function used by the spectral subtraction algorithm is as follows:   First calculate the noise masking threshold of different Bark domains of each frame of speech, and then get the adaptive subtraction parameter according to the noise masking threshold , : If the masking threshold is high, the residual noise will be masked naturally to make it inaudible to the human ear. In this case, the subtraction parameters take their minimum value; when the masking threshold is low, the residual noise will have a great influence on the human ear. large, it is necessary to reduce it; for each frame m, the masking threshold The minimum value of and the subtraction parameter per frame and is related to the maximum value; the application of the subtraction parameter has the following relationship: , (10) in, and respectively The minimum and maximum values of ; , and , are the parameters , The minimum and maximum values of ; when hour, ;when hour, ; where and They are the minimum and maximum values of the masking threshold obtained frame by frame; in the experiment, we set the values of each parameter as follows: v. Real-time noise power spectrum estimation; using a noise power spectrum estimation method based on constrained variance spectrum smoothing and minimum value tracking;  ⅵ. Speech enhancement system; get adaptive subtraction parameters according to the masking threshold , ; The second step: two-dimensional enhancement of speech; 2.1 Two-dimensional noise erosion algorithm The two-dimensional noise erosion algorithm for the speech spectrum is determined by the following process; first, the short-time Fourier transform is performed on the speech, and the spectrum of each frame Calculated by the following formula: (11) is the speech signal of the mth frame, is the frequency spectrum of the m- th frame speech signal; N is the length of the frame and the number of short-time Fourier transform points; Is the Hamming window; the speech signal power spectrum of each frame can be expressed as: (12) That is, it is defined as the spectrum of the speech signal; right The 2D noise erosion of is defined as: (13) in is a structural element, yes domain of definition, yes domain of definition; translation parameters gotta be within the domain of definition, and gotta be within the domain of definition; Structural elements of the two-dimensional noise erosion algorithm for the structural shape of the residual noise spectrum with weak energy is defined as: (14) 2.2 Two-dimensional Speech Expansion Algorithm Results for 2D noise erosion , two-dimensional speech expansion algorithm is defined by: (15) in is a structural element, yes domain of definition, yes domain of definition; Therefore, the structural element in the two-dimensional speech expansion algorithm is defined as the following shape: (16) Step 3: Perceptual Spectral Structure Boundary (PSSB) Parameters and Endpoint Detection Algorithm 3.1 Perceptual Spectral Structure Boundary (PSSB) Parameters The present invention uses the neighborhood model in the formula (17) to approach the result of speech two-dimensional enhancement gradient;                                           (17) is the center point of this neighborhood model; and the gradient of the center neighborhood can be expressed by the following formula: (18) and Determined by formula (19) and formula (20):                                                 (19)     (20) that is The boundary of , which can describe the boundary information of the continuous distribution of the speech signal in the noisy speech spectrum; The perceptual spectral structure boundary parameter PSSB is proposed as follows: (twenty one) in is the PSSB parameter of the mth frame, and M is the total number of frames; 3.2 Voice endpoint detection A detection method aimed at the distribution characteristics of speech continuity is used to distinguish between voiced segments and unvoiced segments at the endpoints; the specific endpoint detection method is as follows: (1) at first detect the speech segment that PSSB parameter is greater than threshold value a and distributes m frame continuously, and this segment is the voiced sound segment that detects; (2) Based on this segment, all segments that are connected with this segment and are continuously greater than or equal to the threshold b are defined as speech segments; the value of the threshold b is relatively small. In the experiment, the value of b is between 0.01 and 0.05. It has a better recognition result; in this way, the unvoiced segment with a smaller PSSB value can be recognized; (3) The start and end points of this speech segment are the speech endpoints. 3. a kind of speech endpoint detection algorithm that adopts perceptual spectrum structure boundary parameter according to claim 2, it is characterized in that: experimental design is under different SNR environments; The low SNR speech of input is 16k sampling, 16 bit quantization. 4. A speech endpoint detection algorithm using perceptual spectrum structure boundary parameters according to claim 2, characterized in that: Hamming window is used, the frame length is 256, and the frame shift is 128. 5. a kind of speech endpoint detection algorithm that adopts perceptual spectrum structure boundary parameter according to claim 2, is characterized in that: speech is selected from TIMIT speech database, and white noise is from NoiseX-92 noise database.