Voice activity modification frame acquiring method, and voice activity detection method and apparatus

Voice Activity Modification Frame Acquiring Method, and Voice Activity Detection Method and Apparatus

Technical Field

The present application relates to, but is not limited to, the field of communications.

Background

In normal voice calls, a user sometimes speaks and sometimes listens. At this time, an inactive speech phase may appear in the call process. In normal cases, a total inactive speech phase of both parties in a call exceeds 50% of a total time length of voice coding of the two parties of the call. In the non-active speech phase, there is only a background noise, and there is generally no useful information in the background noise. With this fact, in the process of voice signal processing, an active speech and a non-active speech are detected through a Voice Activity Detection (VAD for short) algorithm and are processed using different methods respectively. Many voice coding standards, such as Adaptive Multi-Rate (AMR) and Adaptive Multi-Rate Wideband (AMR-WB for short), support the VAD function. In terms of efficiency, the VAD of these encoders cannot achieve good performance under all typical background noises. Especially in an unstable noise, these encoders have low VAD efficiency. For music signals, the VAD sometimes has error detection, resulting in significant quality degradation of the corresponding processing algorithm.

Summary

The following is an overview of the subjects which are described in detail herein. This overview is not intended to limit the embodiments.

The embodiments of the present invention provide a method for acquiring a number of

CA 02990328 2017-12-20 modified frames for active sound and a method and apparatus for voice activity detection (VAD), to solve the problem of low accuracy for the voice activity detection.

The embodiments of the present invention provide a method for acquiring a number of modified frames for active sound, including:

acquiring a voice activity detection, VAD, decision result of a current frame;

acquiring a number of hangover frames for active sound;

acquiring a number of background noise updates; and acquiring the number of modified frames for active sound according to the voice activity detection decision result of the current frame, the number of background noise updates and the number of hangover frames for active sound.

In an exemplary embodiment, acquiring a voice activity detection decision result of a current frame includes:

acquiring a sub-band signal and a spectrum amplitude of the current frame;

calculating a frame energy parameter, a spectral centroid feature and a time-domain stability feature of the current frame according to the sub-band signals; and calculating a spectral flatness feature and a tonality feature according to the spectrum amplitudes;

calculating a signal-to-noise ratio, SNR, parameter of the current frame according to background noise energy estimated from a previous frame, the frame energy parameter and energy of SNR sub-bands of the current frame;

calculating a tonality signal flag of the current frame according to the frame energy parameter, the spectral centroid feature, the time-domain stability feature, the spectral flatness feature, and the tonality feature; and calculating the VAD decision result according to the tonality signal flag, the SNR parameter, the spectral centroid feature, and the frame energy parameter.

In an exemplary embodiment,

CA 02990328 2017-12-20 the frame energy parameter is a weighted cumulative value or a direct cumulative value of energy of various sub-band signals;

the spectral centroid feature is a ratio of a weighted cumulative value and an unweighted cumulative value of the energy of all or a part of the sub-band signals, or is a value obtained by performing smooth filtering on the ratio;

the time-domain stability feature is a desired ratio of a variance of the amplitude cumulative values and a square of the amplitude cumulative values, or is a product of the ratio and a coefficient;

the spectral flatness feature is a ratio of a geometric mean and an arithmetic mean of a predetermined plurality of certain spectrum amplitudes, or is a product of the ratio and a coefficient; and the tonality feature is obtained by calculating a correlation value of intra-frame spectral difference coefficients of two adjacent frame signals, or is obtained by continuing to perform smooth filtering on the correlation value.

In an exemplary embodiment, calculating the voice activity detection decision result according to the tonality signal flag, the SNR parameter, the spectral centroid feature, and the frame energy parameter includes:

acquiring a long-time SNR by computing a ratio of average energy of long-time active frames to average energy of long-time background noise for the previous frame;

acquiring an average total SNR of all sub-bands by calculating an average value of SNR of all sub-bands for a plurality of frames closest to the current frame;

acquiring an SNR threshold for making VAD decision according to the spectral centroid feature, the long-time SNR, the number of continuous active frames and the number of continuous noise frames;

acquiring an initial VAD decision according to the SNR threshold for VAD and the SNR

CA 02990328 2017-12-20 parameter; and acquiring the VAD decision result by updating the initial VAD decision according to the tonality signal flag, the average total SNR of all sub-bands, the spectral centroid feature, and the long-time SNR.

In an exemplary embodiment, acquiring the number of modified frames for active sound according to the voice activity detection decision result of the current frame, the number of background noise updates and the number of hangover frames for active sound includes:

when the VAD decision result indicates the current frame is an active frame and the number of background noise updates is less than a preset threshold, the number of modified frames for active sound is selected as a maximum value of a constant and the number of hangover frames for active sound.

In an exemplary embodiment, obtaining the number of hangover frames for active sound includes:

setting an initial value of the number of hangover frames for active sound.

In an exemplary embodiment, acquiring the number of hangover frames for active sound includes;

acquiring a sub-band signal and a spectrum amplitude of the current frame;

calculating a long-time SNR and an average total SNR of all sub-bands according to the sub-band signal, and obtaining the number of hangover frames for active sound by updating the current number of hangover frames for active sound according to the VAD decision results of a plurality of previous frames, the long-time SNR, the average total SNR of all sub-bands, and the VAD decision result of the current frame.

In an exemplary embodiment, calculating a long-time SNR and an average total SNR of all sub-bands according to the sub-band signal includes:

calculating the long-time SNR through the ratio of the average energy of long-time

CA 02990328 2017-12-20 active frames and the average energy of long-time background noise calculated by using the previous frame of the current frame; and calculating an average value of SNRs of all sub-bands of a plurality of frames closest to the current frame to obtain the average total SNR of all sub-bands.

In an exemplary embodiment, a precondition for modifying the current number of hangover frames for active sound is that a voice activity detection flag indicates that the current frame is an active frame.

In an exemplary embodiment, updating the current number of hangover frames for active sound to acquire the number of hangover frames for active sound includes:

when acquiring the number of hangover frames for active sound, if a number of continuous active frames is less than a set first threshold and the long-time SNR is less than a set threshold, the number of hangover frames for active sound is updated by subtracting the number of continuous active frames from the minimum number of continuous active frames; and if the average total SNR of all sub-bands is greater than a set threshold and the number of continuous active frames is greater than a set second threshold, setting a value of the number of hangover frames for active sound according to the value of the long-time SNR.

In an exemplary embodiment, acquiring a number of background noise updates includes:

acquiring a background noise update flag; and calculating the number of background noise updates according to the background noise update flag.

In an exemplary embodiment, calculating the number of background noise updates according to the background noise update flag includes:

setting an initial value of the number of background noise updates.

In an exemplary embodiment, calculating the number of background noise updates according to the background noise update flag includes:

CA 02990328 2017-12-20 when the background noise update flag indicates that a current frame is a background noise and the number of background noise updates is less than a set threshold, adding the number of background noise updates by 1.

In an exemplary embodiment, acquiring a background noise update flag includes:

acquiring a sub-band signal and a spectrum amplitude of the current frame;

calculating a frame energy parameter, a spectral centroid feature and a time-domain stability feature according to the sub-band signal; and calculating a spectral flatness feature and a tonality feature according to the spectrum amplitude; and performing background noise detection according to the spectral centroid feature, the time-domain stability feature, the spectral flatness feature, the tonality feature, and the frame energy parameter to acquire the background noise update flag.

In an exemplary embodiment, the frame energy parameter is a weighted cumulative value or a direct cumulative value of energy of various sub-band signals;

the spectral centroid feature is a ratio of a weighted cumulative value and an unweighted cumulative value of the energy of ail or a part of the sub-band signals, or is a value obtained by performing smooth filtering on the ratio;

the time-domain stability feature is a desired ratio of a variance of the frame energy amplitudes and a square of the amplitude cumulative values, or is a product of the ratio and a coefficient; and the spectral flatness parameter is a ratio of a geometric mean and an arithmetic mean of a predetermined plurality of spectrum amplitudes, or is a product of the ratio and a coefficient.

In an exemplary embodiment, performing background noise detection according to the spectral centroid feature, the time-domain stability feature, the spectral flatness feature, the tonality feature, and the frame energy parameter to acquire the background noise update flag

CA 02990328 2017-12-20 includes:

setting the background noise update flag as a first preset value;

determining that the current frame is not a noise signal and setting the background noise update flag as a second preset value if any of the following conditions is true:

the time-domain stability feature is greater than a set threshold;

a smooth filtered value of the spectral centroid feature value is greater than a set threshold, and a value of the time-domain stability feature is also greater than a set threshold;

a value of the tonality feature or a smooth filtered value of the tonality feature is greater than a set threshold, and a value of the time-domain stability feature is greater than a set threshold;

a value of a spectral flatness feature of each sub-band or a smooth filtered value of the spectral flatness feature of each sub-band is less than a respective corresponding set threshold; or a value of the frame energy parameter is greater than a set threshold.

The embodiments of the present invention provide a method for voice activity detection, including:

acquiring a first voice activity detection decision result;

acquiring a number of hangover frames for active sound;

acquiring a number of background noise updates;

calculating a number of modified frames for active sound according to the first voice activity detection decision result, the number of background noise updates, and the number of hangover frames for active sound;

acquiring a second voice activity detection decision result; and calculating the voice activity detection decision result according to the number of

CA 02990328 2017-12-20 modified frames for active sound and the second voice activity detection decision result.

In an exemplary embodiment, calculating the voice activity detection decision result according to the number of modified frames for active sound and the second voice activity detection decision result includes:

when the second voice activity detection decision result indicates that the current frame is an inactive frame and the number of modified frames for active sound is greater than 0, setting the voice activity detection decision result as an active frame, and reducing the number of modified frames for active sound by 1.

In an exemplary embodiment, acquiring a first voice activity detection decision result includes:

acquiring a sub-band signal and a spectrum amplitude of a current frame;

calculating a frame energy parameter, a spectral centroid feature and a time-domain stability feature of the current frame according to the sub-band signal; and calculating a spectral flatness feature and a tonality feature according to the spectrum amplitude;

calculating a signal-to-noise ratio parameter of the current frame according to background noise energy acquired from a previous frame, the frame energy parameter and signal-to-noise ratio sub-band energy;

calculating a tonality signal flag of the current frame according to the frame energy parameter, the spectral centroid feature, the time-domain stability feature, the spectral flatness feature, and the tonality feature; and calculating the first voice activity detection decision result according to the tonality signal flag, the signal-to-noise ratio parameter, the spectral centroid feature, and the frame energy parameter.

In an exemplary embodiment, the frame energy parameter is a weighted cumulative value or a direct cumulative value of energy of various sub-band signals;

CA 02990328 2017-12-20 the spectral centroid feature is a ratio of a weighted cumulative value and an unweighted cumulative value of the energy' of all or a part of the sub-band signals, or is a value obtained by performing smooth filtering on the ratio;

the time-domain stability feature is a desired ratio of a variance of the amplitude cumulative values and a square of the amplitude cumulative values, or is a product of the ratio and a coefficient;

the spectral flatness feature is a ratio of a geometric mean and an arithmetic mean of a predetermined plurality of spectrum amplitudes, or is a product of the ratio and a coefficient; and the tonality' feature is obtained by calculating a correlation value of intra-frame spectral difference coefficients of two adjacent frame signals, or is obtained by continuing to perform smooth filtering on the correlation value.

In an exemplary embodiment, calculating the first voice activity detection decision result according to the tonality signal flag, the signal-to-noise ratio parameter, the spectral centroid feature, and the frame energy parameter includes:

calculating a long-time SNR through a ratio of average energy of long-time active frames and average energy of long-time background noise calculated at the previous frame;

calculating an average value of SNRs of all sub-bands of a plurality of frames closest to the current frame to acquire an average total SNR of all sub-bands;

acquiring a voice activity detection decision threshold according to the spectral centroid feature, the long-time SNR, the number of continuous active frames and the number of continuous noise frames;

calculating an initial voice activity detection decision result according to the voice activity detection decision threshold and the signal-to-noise ratio parameter; and modifying the initial voice activity detection decision result according to the tonality

CA 02990328 2017-12-20 signal flag, the average total SNR of all sub-bands, the spectral centroid feature, and the long-time SNR to acquire the first voice activity detection decision result.

In an exemplary embodiment, obtaining the number of hangover frames for active sound includes:

setting an initial value of the number of hangover frames for active sound.

In an exemplary embodiment, acquiring the number of hangover frames for active sound includes:

acquiring a sub-band signal and a spectrum amplitude of a current frame; and calculating a long-time SNR and an average total SNR of all sub-bands according to the sub-band signals, and modifying the current number of hangover frames for active sound according to voice activity detection decision results of a plurality of previous frames, the long-time SNR, the average total SNR of all sub-bands, and the first voice activity detection decision result.

In an exemplary embodiment, calculating a long-time SNR and an average total SNR of all sub-bands according to the sub-band signal includes:

calculating the long-time SNR through the ratio of the average energy of long-time active frames and the average energy of long-time background noise calculated by using the previous frame of the current frame; and calculating an average value of SNRs of all sub-bands of a plurality of frames closest to the current frame to acquire the average total SNR of all sub-bands.

In an exemplary embodiment, a precondition for correcting the current number of hangover frames for active sound is that a voice activity flag indicates that the current frame is an active frame.

In an exemplary embodiment, modifying the current number of hangover frames for active sound includes:

CA 02990328 2017-12-20 if the number of continuous voice frames is less than a set first threshold and the long-time SNR is less than a set threshold, the number of hangover frames for active sound being equal to a minimum number of continuous active frames minus the number of continuous active frames; and if the average total SNR of all sub-bands is greater than a set second threshold and the number of continuous active frames is greater than a set threshold, setting a value of the number of hangover frames for active sound according to a size of the long-time SNR.

In an exemplary embodiment, acquiring a number of background noise updates includes:

acquiring a background noise update flag; and calculating the number of background noise updates according to the background noise update flag.

In an exemplary embodiment, calculating the number of background noise updates according to the background noise update flag includes:

setting an initial value of the number of background noise updates.

In an exemplary embodiment, calculating the number of background noise updates according to the background noise update flag includes:

when the background noise update flag indicates that a current frame is a background noise and the number of background noise updates is less than a set threshold, adding the number of background noise updates by 1.

In an exemplary embodiment, acquiring a background noise update flag includes:

acquiring a sub-band signal and a spectrum amplitude of a current frame;

calculating values of a frame energy parameter, a spectral centroid feature and a time-domain stability feature according to the sub-band signal; and calculating values of a spectral flatness feature and a tonality feature according to the spectrum amplitude; and performing background noise detection according to the spectral centroid feature, the

CA 02990328 2017-12-20 time-domain stability feature, the spectral flatness feature, the tonality feature, and the frame energy parameter to acquire the background noise update flag.

In an exemplary embodiment, the frame energy parameter is a weighted cumulative value or a direct cumulative value of energy of various sub-band signals;

the spectral centroid feature is a ratio of a weighted cumulative value and an unweighted cumulative value of the energy of all or a part of the sub-band signals, or is a value obtained by performing smooth filtering on the ratio;

the time-domain stability feature is a desired ratio of a variance of the frame energy amplitudes and a square of the amplitude cumulative values, or is a product of the ratio and a coefficient; and the spectral flatness parameter is a ratio of a geometric mean and an arithmetic mean of a predetermined plurality of spectrum amplitudes, or is a product of the ratio and a coefficient.

In an exemplary embodiment, performing background noise detection according to the spectral centroid feature, the time-domain stability feature, the spectral flatness feature, the tonality feature, and the frame energy parameter to acquire the background noise update flag includes:

setting the background noise update flag as a first preset value;

determining that the current frame is not a noise signal and setting the background noise update flag as a second preset value if any of the following conditions is true:

the time-domain stability feature is greater than a set threshold;

a smooth filtered value of the spectral centroid feature value is greater than a set threshold, and a value of the time-domain stability feature is also greater than a set threshold;

a value of the tonality feature or a smooth filtered value of the tonality feature is greater than a set threshold, and a value of the time-domain stability feature is greater than a set threshold;

CA 02990328 2017-12-20 a value of a spectral flatness feature of each sub-band or a smooth filtered value of the spectral flatness feature of each sub-band is less than a respective corresponding set threshold;

or a value of the frame energy parameter is greater than a set threshold.

In an exemplary embodiment, calculating the number of modified frames for active sound according to the first voice activity detection decision result, the number of background noise updates and the number of hangover frames for active sound includes:

when the first voice activity detection decision result is an active frame and the number of background noise updates is less than a preset threshold, the number of modified frames for active sound being a maximum value of a constant and the number of hangover frames for active sound.

The embodiments of the present invention provide an apparatus for acquiring a number of modified frames for active sound, including:

a first acquisition unit arranged to acquire a voice activity detection decision result of a current frame;

a second acquisition unit arranged to acquire a number of hangover frames for active sound;

a third acquisition unit arranged to acquire a number of background noise updates; and a fourth acquisition unit arranged to acquire the number of modified frames for active sound according to the voice activity detection decision result of the current frame, the number of background noise updates and the number of hangover frames for active sound.

The embodiments of the present invention provide an apparatus for voice activity detection, including:

a fifth acquisition unit arranged to acquire a first voice activity detection decision result;

a sixth acquisition unit arranged to acquire a number of hangover frames for active

CA 02990328 2017-12-20 sound;

a seventh acquisition unit arranged to acquire a number of background noise updates;

a first calculation unit arranged to calculate a number of modified frames for active sound according to the first voice activity detection decision result, the number of background noise updates, and the number of hangover frames for active sound;

an eighth acquisition unit arranged to acquire a second voice activity detection decision result; and a second calculation unit arranged to calculate the voice activity detection decision result according to the number of modified frames for active sound and the second voice activity detection decision result.

A computer readable storage medium, has computer executable instructions stored thereon for performing any of the methods as described above.

The embodiments of the present invention provide a method for acquiring a number of modified frames for active sound, and a method and apparatus for voice activity detection. Firstly, a first voice activity detection decision result is obtained, a number of hangover frames for active sound is obtained, and a number of background noise updates is obtained, and then a number of modified frames for active sound is calculated according to the first voice activity detection decision result, the number of background noise updates and the number of hangover frames for active sound, and a second voice activity detection decision result is obtained, and finally, the voice activity detection decision result is calculated according to the number of modified frames for active sound and the second voice activity detection decision result, which can improve the detection accuracy of the VAD.

After reading and understanding the accompanying drawings and detailed description, other aspects can be understood.

CA 02990328 2017-12-20

Brief Description of Drawings

Fig. 1 is a flowchart of a method for voice activity detection according to embodiment one of the present invention;

Fig. 2 is a diagram of a process of obtaining a VAD decision result according to the embodiment one of the present invention;

Fig. 3 is a flowchart of a method for detecting a background noise according to embodiment two of the present invention;

Fig. 4 is a flowchart of a method for correcting the current number of hangover frames for active sound in VAD decision according to embodiment three of the present invention;

Fig. 5 is a flowchart of a method for acquiring the number of modified frames for active sound according to embodiment four of the present invention;

Fig. 6 is a structural diagram of an apparatus for acquiring the number of modified frames for active sound according to the embodiment four of the present invention;

Fig. 7 is a flowchart of a method for voice activity detection according to embodiment five of the present invention; and

Fig. 8 is a structural diagram of an apparatus for voice activity detection according to the embodiment five of the present invention.

Detailed Description

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It is to be illustrated that the embodiments in the present application and features in the embodiments can be combined with each other randomly without conflict.

The steps shown in the flowchart of the accompanying drawings can be performed in a computer system such as a set of computer-executable instructions. Further, although a logical

CA 02990328 2017-12-20 order is shown in the flowchart, in some cases, the steps shown or described can be performed in an order different from that described here.

Symbol description: without special description, in the following embodiments, a right superscript [i] represents a frame serial number, [0] represents a current frame, and [-1] represents a previous frame. For example, and represent smoothed spectrums of the current frame and the previous frame.

Embodiment one

The embodiment of the present invention provides a method for voice activity detection, as shown in Fig. 1, including the following steps.

In step 101, a sub-band signal and a spectrum amplitude of a current frame are obtained.

The present embodiment is described by taking an audio stream with a frame length of 20ms and a sampling rate of 32kHz as an example. In conditions with other frame lengths and sampling rates, the method herein is also applicable.

A time domain signal of the current frame is input into a filter bank, to perform sub-band filtering calculation so as to obtain a sub-band signal of the filter bank;

In this embodiment, a 40-channel filter bank is used, and the method herein is also applicable to filter banks with other numbers of channels. It is assumed that the input audio signal is s_HP(n), L_c is 40, which is the number of channels of the filter bank, w_c is a window function with a window length of 10 L_c , and the sub-band signal is

X(k,l)~ X_CR (I, k)+i-X_a(J,k), herein X_CR and X_CI are the real and the imaginary parts of the sub-band signal. A calculation method for the sub-band signal is as follows:

<img file="CA2990328C_D0001.tif" />

-WcOOÂc-wi'^ClOLc-M+Z'Zcicos

<img file="CA2990328C_D0002.tif" />

<img file="CA2990328C_D0003.tif" />

<img file="CA2990328C_D0004.tif" />

<img file="CA2990328C_D0005.tif" />

-w_c(10Zc-M)-j_WF(10Zc-zi + Z-Z._c)sin —[«+—+—

Lq \ 2 2 j k+2

CA 02990328 2017-12-20

Herein I is a sub-band time index and 0&lt;/&lt;15, k is a sub-band index with 0&lt;k&lt;L_c.

Time-frequency transform is performed on the sub-band signal of the filter bank and the spectrum amplitude is calculated.

Herein, the embodiment of the present invention can be realized by performing time-frequency transform on all the sub-bands of the filter bank or a part of the sub-bands of the filter bank, and calculating the spectrum amplitude. The time-frequency transform method according to the embodiment of the present invention may be Discrete Fourier Transform (DFT for short), Fast Fourier Transformation (FFT for short), Discrete Cosine Transform (DCT for short) or Discrete Sine Transform (DST for short). This embodiment uses the DFT as an example to illustrate its implementation method. A calculation process is as follows.

A 16-point DFT transform is performed on data of 16 time sample points on each of the sub-band of the filter bank with indexes of 0 to 9, to further improve the spectral resolution and calculate the amplitude at each frequency point so as to obtain the spectrum amplitude

A calculation equation of the time-frequency transform is as follows:

XdftWJ]=¹⁶ < ¹⁰>° J < 16.

1=0

A calculation process of the amplitude at each frequency point is as follows.

Firstly, energy of an array X_DFT[k,j] at each point is calculated, and a calculation equation as follows.

XoFrj’OH'U¹’j] = ((Re(^R, j]))² + (Im(A^> J]))²);O < k < 10,0 < j < 16

Herein, Re and Im represent the real and the imaginary parts of the spectral coefficients X_DFT respectively.

If k is an even number, the following equation is used to calculate the spectrum

CA 02990328 2017-12-20 amplitude at each frequency point:

^spW + j) = JXDFT_POwi^k,7] T XOFT_P0WW5 - 7], 0 &lt; k &lt; 10,0 £ J &lt; 8

If k is an odd number, the following equation is used to calculate the spectrum amplitude at each frequency point:

^sp+ 7 - j) = X_Drr pow[Λ,/]+X dit _pon4.^k 45 ^— /]> 0 i < 10,0 < 7 < 8

A is the spectrum amplitude after the time-frequency transform is performed.

In step 102, The frame energy features, spectral centroid features and time-domain stability features of the current frame are calculated according to the sub-band signal, and the spectral flatness features and tonality features arc calculated according to the spectrum amplitude.

Herein the frame energy parameter is a weighted cumulative value or a direct cumulative value of energy of all sub-band signals, herein:

a) energy £_iA[k] = Ec(k) of each sub-band of the filter bank is calculated according to the sub-band signal X[k,l] of the filter bank:

0&lt;k&lt;L_c

Herein, E_c{t,k)=(X_CR(t,k)Ÿ ^X_cl(t,k))² 0&lt;/&lt;15, 0&lt;k&lt;L_c.

b) Energy of a part of the sub-bands of the filter bank which are acoustically sensitive or energy of all the sub-bands of the filter bank are cumulated to obtain the frame energy parameter.

Herein, according to a psychoacoustic model, the human ear is less sensitive to sound at very low frequencies (for example, below 100Hz) and high frequencies (for example, above 20kHz). For example, in the embodiment of the present invention, it is considered that among the sub-bands of the filter bank which are arranged in an order from low frequency to high frequency, the second sub-band to the last but one sub-band are primary sub-bands of the filter bank which are acoustically sensitive, energy of a part or all of the sub-bands of the

CA 02990328 2017-12-20 filter bank which are acoustically sensitive is cumulated to obtain a frame energy parameter 1, and a calculation equation is as follows:

esbend e„- '£ M*) n»c_sb _$lart

Herein, ^e _sb start j_{s a start} sub-band index, with a value range of [0,6]. 6 sb 6 fid

-- is an end sub-band index, with a value greater than 6 and less than the total number of the sub-bands.

A value of the frame energy parameter 1 is added to a weighted value of the energy of a part or all of sub-bands of the filter bank which are not used when the frame energy parameter 1 is calculated to obtain a frame energy parameter 2, and a calculation equation thereof is as follows:

e _$b _stan-\ num-band

E_a = E_a+e _ scale) £ Ec(k)+e _ scale! · £ Ec(k).

η·0 n*e_sb_end+l herein e_ scale) and e_ scale) are weighted scale factors with value ranges of [0,1 ] respectively, and ^num _ band is the total number of sub-bands.

The spectral centroid features are the ratio of the weighted sum to the non-weighted sum of energies of all sub-bands or partial sub-bands, herein:

The spectral centroid features are calculated according to the energies of sub-bands of the filter bank. A spectral centroid feature is the ratio of the weighted sum to the non-weighted sum of energies of all or partial sub-bands, or the value is obtained by applying smooth filtering to this ratio.

The spectral centroid features can be obtained by the following sub-steps:

a: A sub-band division for calculating the spectral centroid features is as follows.

<td> Spectral centroid</td><td> Spectral centroid feature</td><td> Spectral centroid feature</td>

CA 02990328 2017-12-20

<td> feature number</td><td> start sub-band index ^SP^C _ ^siart _ band</td><td> end sub-band index ^SP^C - ^end _ band</td>

<td> 0</td><td> 0</td><td> 10</td>

<td> 1</td><td> 1</td><td> 24</td>

b: values of two spectral centroid features, which are a first interval spectral centroid feature and a second interval spectral centroid feature, are calculated using the interval division manner for the calculation of the spectral centroid feature in a and the following equation.

spc _benii(k)-spc _sum jKind(k) £ (« + !)· [«+spc _ start _ band(k)] + Delta\ sp _center[k\ = —---------------------------------’° * ^k *

E^ln* spc _ start _ band (A )]+Delta2 n=0

Delta) and Delta! are small offset values respectively, with a value range of (0,1). Herein, is a spectral centroid feature number.

c: a smooth filtering operation is performed on the first interval spectral centroid feature sp _center[Q] to obtain a smoothed value of the spectral centroid feature, that is, a smooth filtered value of the first interval spectral centroid feature, and a calculation process is as follows:

sp _ center[2] = sp _ center^ [2] · spc _sm_ scale + sp_ center[0] · (1 - spc _sm_ scale)

Herein, spc _sm_scale is a smooth filtering scale factor of the spectral centroid feature, and sp~center__}[2] represents the smoothed value of spectral centroid feature in the previous frame, with an initial value of 1.6.

The time-domain stability feature is the ratio of the variance of the sums of amplitudes to the expectation of the squaredamplitudes, or the ratio multiplied by a factor, herein:

The time-domain stability features are computed with the energy features of the most recent several frames. In the present embodiment, the time-domain stability feature is

CA 02990328 2017-12-20 calculated using frame energies of 40 latest frames. The calculation steps arc as follows.

Firstly, the energy amplitudes of the 40 latest frame signals are calculated, and a calculation equation is as follows:

Amp^n] = ^Ε_ι2(η)+e _ offset',Q n &lt; 40;

Herein, e _ offset is an offset value, with a value range of [0, 0.1].

Next, by adding together the energy amplitudes of two adjacent frames from the current frame to the 40^th previous frame, 20 sums of energy amplitudes are obtained. A calculation equation is as follows:

Amp^ri) - Amp_n(-2n) + Amp„(-2n-l);0 &lt; n &lt; 20;

Herein, when m=0, Amp_a represents an energy amplitude of the current frame, and when »&lt;0, Amp_t[ represents the energy amplitude of the η^Λ previous frame from the current frame.

Finally, the time-domain stability feature ltd_stable_rateO is obtained by calculating the ratio of the variance to the average energy of the 20 sums of amplitudes closet to the current frame. A calculation equation is as follows:

/ ] .9 λ²

Σ ^ΑηΡα(η)~™Σ^Α,ηΡΜ ltd_stable_rateO = ---------—--------— ;

Amp_n(ri)² + Delta n=0

Spectral flatness feature is the ratio of the geometric mean to the arithmetic mean of certain spectrum amplitude, or the ratio multiplied by a factor.

The spectrum amplitude is smoothed to obtain:

A™(i) = 0.7^(/)+0.3^(0, 0 &lt; i &lt; N_a

Herein, 4^J(?j and 4fJ‘(0 represent the smoothed spectrum amplitudes of the current frame and the previous frame, respectively, and N_A is the number of the spectrum

CA 02990328 2017-12-20 amplitudes.

It is to be illustrated that the predetermined several spectrum amplitudes described in the embodiment of the present invention may be partial of the spectrum amplitudes selected according to the experience of those skilled in the art or may also be a part of the spectrum amplitudes selected according to practical situations.

In this embodiment, the spectrum amplitude is divided into three frequency regions, and the spectral flatness features are computed for these frequency regions. A division manner thereof is as follows.

Sub-band division for computing spectral flatness features

<td> Spectral flatness number (k)</td><td></td><td></td>

<td> 0</td><td> 5</td><td> 19</td>

<td> 1</td><td> 20</td><td> 39</td>

<td> 2</td><td> 40</td><td> 64</td>

Let N(k) = N_{À Sfarl}(k) represent the number of spectrum amplitudes used to calculate the spectral flatness features, F^k) :

/7 7 /A — * ” Wt)·*' ^{SF S!P} '

Finally, the spectral flatness features of the current frame are smoothed to obtain final spectral flatness features of the current frame:

¢(4)=0.85^(4)+0.15^(4).

Herein, F™(k) and are the smoothed spectral flatness features of the current

CA 02990328 2017-12-20 frame and the previous frame, respectively.

The tonality features are obtained by computing the correlation coefficient of the intra-frame spectrum amplitude difference of two adjacent frames, or obtained by further smoothing the correlation coefficient.

A calculation method of the correlation coefficient of the intra-frame spectrum amplitude differences of the two adjacent frame signals is as follows.

The tonality feature is calculated according to the spectrum amplitude, herein the tonality feature may be calculated according to all the spectrum amplitudes or a part of the spectrum amplitudes.

The calculation steps are as follows.

a): Spectrum-amplitude differences of two adjacent spectrum amplitudes are computed for partial (not less than 8 spectrum amplitudes) or all spectrum amplitudes in the current frame. If the difference is smaller than 0, set it to 0, and a group of non-negative spectrum-amplitude differences is obtained:

0^(0 = if A_sp(i+6) &lt;ΛΛ+⁵)

A_sp (/+6)- Α_ψ (i+5) otherwise

b): The correlation coefficient between the non-negative spectrum-amplitude differences of the current frame obtained in Step a) and the non-negative spectrum-amplitude differences of the previous frame is computed to obtain the first tonality feature as follows:

N

TR ~

<img file="CA2990328C_D0006.tif" />

Herein, ^sp is a non-negative spectrum-amplitude difference of the previous frame.

c): The first tonality feature is smoothed to obtain the value of second tonality feature

CA 02990328 2017-12-20 and the third tonality feature ^(2), herein a comer mark 0 represents the current frame, and the calculation equation is as follows:

/y(0) = F^

F/^0|(l) = 0.96F/(1)+0.04F_TR .

F/^0,(2) = 0.90 F/ (2) + 0.10F_re

In step 103, signal-to-noise ratio (SNR) parameters of the current frame are calculated according to background noise energy estimated from the previous frame, the frame energy parameter and the energy of signal-to-noise ratio sub-bands of the current frame.

The background noise energy of the previous frame may be obtained using an existing method.

If the current frame is a start frame, a default initial value is used as background noise energy of SNR sub-bands. In principle, the estimation of background noise energy of the SNR sub-bands of the previous frame is the same as that of the current frame. The estimation of the background energy of SNR sub-bands of the current frame can be known with reference to step 107 of the present embodiment. Herein, the SNR parameters of the current frame can be achieved using an existing method. Alternatively, the following method is used:

Firstly, the sub-bands of the filter bank arc re-divided into a plurality of SNR sub-bands, and division indexes are as follows in the following table.

<td> SNR sub-band serial number</td><td> Start filter bank sub-band serial number ( Sub_Start_index )</td><td> End filter bank sub-band serial number (Sub_end_index)</td>

<td> 0</td><td> 0</td><td> 0</td>

<td> 1</td><td> 1</td><td> 1</td>

<td> 2</td><td> 2</td><td> 2</td>

CA 02990328 2017-12-20

<td> 3</td><td> 3</td><td> 3</td>

<td> 4</td><td> 4</td><td> 4</td>

<td> 5</td><td> 5</td><td> 5</td>

<td> 6</td><td> 6</td><td> 6</td>

<td> 7</td><td> 7</td><td> 8</td>

<td> 8</td><td> 9</td><td> 10</td>

<td> 9</td><td> 11</td><td> 12</td>

<td> 10</td><td> 13</td><td> 16</td>

<td> 11</td><td> 17</td><td> 24</td>

<td> 12</td><td> 25</td><td> 36</td>

Secondly, each SNR sub-band energy of the current frame is calculated according to the division manner of the SNR sub-bands. The calculation equation is as follows:

Sub end_index(w)

Σ ^[ΠΟ&lt;«&lt;13;

t=Sub _Start ^mdex(rt)

Then, a sub-band average SNR, SNR1 is calculated according to each SNR sub-band energy of the current frame and each SNR sub-band background noise energy of the previous frame. A calculation equation is as follows:

num band·} rp

SNR\--!---- y log, num band E_{tb2 bg}(n)

E

Herein, is the estimated SNR sub-band background noise energy of the previous frame, and num _band _num(,_{er o}f sub-bands. The principle of obtaining the

SNR sub-band background noise energy of the previous frame is the same as that of obtaining the SNR sub-band background noise energy of the current frame. The process of obtaining the SNR sub-band background noise energy of the current frame can be known with reference to

CA 02990328 2017-12-20 step 107 in the embodiment one below.

Finally, the SNR of all sub-bands, SNR2, is calculated according to estimated energy of background noise over all sub-bands in the previous frame and the frame energy of the current frame;

F

SW2 = log₂-^E

Herein, '-^bg is the estimated energy of background noise over all sub-bands of the previous frame, and the principle of obtaining the energy of background noise over all sub-bands of the previous frame is the same as that of obtaining the energy of background noise over all sub-bands of the current frame. The process of obtaining the energy of background noise over all sub-bands of the current frame can be known with reference to step 107 in the embodiment one below.

In this embodiment, the SNR parameters include the sub-band average SNR, SNRl, and the SNR of all sub-bands, SNR2. The energy of background noise over all sub-bands and each sub-band background noise energy are collectively refereed to as background noise energy.

In step 104, a tonality signal flag of the current frame is calculated according to the frame energy parameter, the spectral centroid feature, the time-domain stability feature, the spectral flatness feature and the tonality feature of the current frame.

In 104a, it is assumed that the current frame signal is a non-tonal signal and a tonal frame flag tonality_ frame is used to indicate whether the current frame is a tonal frame.

In this embodiment, a value of tonality_ frame being 1 represents that the current frame is a tonal frame, and the value being 0 represents that the current frame is a non-tonal frame.

In 104b, it is judged whether the tonality feature or its smoothed value is greater than the corresponding set threshold lomlily_decision_th_rl _or tonality_Λώϊοη_thr2 _{Md if}

CA 02990328 2017-12-20 one of the above conditions is met, step 104c is executed; otherwise, step 104d is executed.

Herein, a value range of ^ίοηα^-decision_thrl _is θ-η, _anj _{a va}]_{ue range} θ£ tonality _ratel _{is [0}.7, 0.99],

In 104c, if the time-domain stability feature lt_stable_rateO is less than a set threshold It_stable-decision_thr\, the spectral centroid feature sp _center[]] is greater than a set threshold spe-decision _thrl, and a spectral flatness feature of any sub-band is less than a respective corresponding preset threshold, it is determined that the current frame is a tonal frame, and the value of the tonal frame flag tonality_ frame is set to 1; otherwise it is determined that the current frame is a non-tonal frame, the value of the tonal frame flag tonality_ frame is set to 0, and step 104d continues to be performed.

Herein, a value range of the threshold Stable-decision_thrl ._g 0.25], and a value range of W-decision _thrl _is[10,1.8].

In 104d, the tonal level feature tonality_d^ree ;_{s U}pd_at_ecj according to the tonal frame flag tonality_ frame. The initial value of tonal level feature ^a^-degree region [0, 1] when the active-sound detection begins. In different cases, calculation methods for the tonal level feature t°^nahty_degree are different.

If the current tonal frame flag indicates that the current frame is a tonal frame, the following equation is used to update the tonal level feature tonality_degrcc :

tonality_degree = tonalitydegree^ · td_scale_A + td_scale_B;

Herein, tonality_degree_₁ is the tonal level feature of the previous frame, with a value range of the initial value thereof of [0,1], td_scale_A is an attenuation coefficient, with a value range of [0,1] and ^-ScaleB j_{s a} cumulative coefficient, with a value range of [0,1]·

In 104e, whether the current frame is a tonal signal is determined according to the updated tonal level feature tonality degree, and the value of the tonality signal flag tonality _flag is set.

CA 02990328 2017-12-20

If the tonal level feature tonality_degree is greater than a set threshold, it is determined that the current frame is a tonal signal; otherwise, it is determined that the current frame is a non-tonal signal.

In step 105, a VAD decision result is calculated according to the tonality signal flag, the SNR parameter, the spectral centroid feature, and the frame energy parameter, and as shown in Fig. 2, the steps are as follows.

In step 105a, the long-time SNR It _snr is obtained by computing the ratio of the average energy of long-time active frames to the average energy of long-time background noise for the previous frame.

The average energy of long-time active frames E^ and the average energy of long-time background noise E_bg are calculated and defined in step 105g. A long-time SNR It _snr is calculated as follows:

E i

It tsnr = log|₀ , herein, in this equation, the long-time SNR ^lt-^snr is expressed logarithmically.

In step 105b, an average value of SNR of all sub-bands SNR2 for multiple frames closest to the current frame is calculated to obtain an average total SNR of all sub-bands SNR2_lt _ave.

A calculation equation is as follows:

F num SNR2lt_ave---:-- V SNR2(n) F _num

SNR2(n) _rep_resent_{s a} value of SNR of all sub-bands SNR2 at the n^th previous frame of the current frame, and F_num is the total number of frames for the calculation of the average value, with a value range of [8,64],

In step 105c, a SNR threshold for making VAD decision snr_thr is obtained according to

CA 02990328 2017-12-20 the spectral centroid feature, the long-time SNR lt_snr, the number of continuous active frames continuous_speech_num, the number of continuous noise frames continuous_noise_num.

Implementation steps are as follows.

Firstly, an initial SNR threshold snr_thr , with a range of [0.1, 2], is set to for example 1.06.

Secondly, the SNR threshold snr_thr is adjusted for the first time according to the spectral centroid feature. The steps are as follows. If the value of the spectral centroid feature sp_center[2] gj._eater than _{a set} threshold ^sP^c_^vad_dec_thrl, then snrjhr is added with an offset value, and in this example, the offset value is taken as 0.05; otherwise, if is greater than ^sP^c_^vad_dec_thr2. j_{s w}jth an offset value, and in this example, the offset value is taken as 0.10; otherwise, snr_thr is added with an offset value, and in this example, the offset value is taken as 0.40, herein value ranges of the thresholds spc_vad_dec_thrl spc_vad_dec_thr2 _are μ 2 2 5]

Then, snr_thr is adjusted for the second time according to the number of continuous active frames continuous_speech_num, the number of continuous noise frames continuous noise num. the average total SNR of all sub-bands SNR2_lt_ave and the long-time SNR It _snr. If the number of continuous active frames continuous_speech_num is greater than a set threshold cpn_vad_dec_thrl, 0.2 is subtracted from snr_thr; otherwise, if the number of continuous noise frames continuous_noise_num is greater than a set threshold cpn_vad_dec_thr2, and SNR2_lt_ave is greater than an offset value plus the long-time SNR It snr multiplied with a coefficient It tsnr_scale, snr_thr is added with an offset value, and in this example, the offset value is taken as 0.1; otherwise, if continuous_noisejium is greater than a set threshold cpn_vad_dec_thr3, snr_thr is added with an offset value, and in this example, the offset value is taken as 0.2; otherwise if continuous_noise_num is greater than a set threshold cpn_vad_dec_thr4, snr_thr is added

CA 02990328 2017-12-20 with an offset value, and in this example, the offset value is taken as 0.1. Herein, value ranges of the thresholds cpn_vad_dec_thrl , cpn_vad_dec_thr2 , cpn_vad_dec_thr3 and cpn_vad_dec_thr4 are [2, 500], and a value range of the coefficient It _tsnr _scale is [0,2]. The embodiment of the present invention can also be implemented by skipping the present step to directly proceed to the final step.

Finally, a final adjustment is performed on the SNR threshold snr thr according to the long-time SNR Itsnr to obtain the SNR threshold snr_thr of the current frame.

The adjustment equation is as follows:.

snr_thr = snr_thr + (It _tsnr -thr _ offset) · thr _ scale

Herein, thr _offset is an offset value, with a value range of [0.5, 3]; and thr_scale is a gain coefficient, with a value range of [0.1,1].

In step 105d, an initial VAD decision is calculated according to the SNR threshold snr thr and the SNR parameters SNRl _and SNR2 calculated at the current frame.

A calculation process is as follows.

If S/VK1 is greater than the SNR threshold snr_thr, it is determined that the current frame is an active frame, and a value of VAD flag vad flag is used to indicate whether the current frame is an active frame. In the present embodiment, 1 is used to represent that the current frame is an active frame, and 0 is used to represent that the current frame is a nonactive frame. Otherwise, it is determined that the current frame is a non-active frame and the value of the VAD flag vad flag is set to 0.

If SNR2 is greater than a set threshold snr2_thr, it is determined that the current frame is an active frame and the value of the VAD flag vad_flag is set to 1. Herein, a value range of snr2_thr is [1.2,5.0].

In step 105e, the initial VAD decision is modified according to the tonality signal flag, the average total SNR of all sub-bands SNR2 lt _ave, the spectral centroid feature, and the

CA 02990328 2017-12-20 long-time SNR ^lt

Steps are as follows.

If the tonality signal flag indicates that the current frame is a tonal signal, that is, tonality flag is 1, then it is determined that the current frame is an active signal and the flag vadflag is set to 1.

If the average total SNR of all sub-bands 5NR2_lt_ave j_s g_rcat_er than _{a se}t threshold SNR2_lt_ave_t_thr\ p|_us i_ong-tj_me snr lt_snr multiplied with a coefficient It tsnr_tscale, _th_en j_t ._g determined that the current frame is an active frame and the flag vad flag is set to 1.

Herein, in the present embodiment, a value range of 5NR2 _lt _ave __thrl -_{g an}j a value range of _^tsnr _tscale -_g

If the average total SNR of all sub-bands SNR2_lt_ave j_s g_reater than _{a set} threshold SNR2 _lt _ave _t_thr2 _{the spectra}i _centroid feature sp_c*nter\2\ j_s greater than a set threshold ^SP _^center anj the long-time SNR It_snr is less than a set threshold It tsnr _t thrl, js determined that the current frame is an active frame, and the flag vad flag is set to 1. Herein, a value range of SNR2 _lt _ave _t _thr2 js μ q, 2.5], a value range of ^SP_^center js [2.0, 4.0], and a value range of It _tsnr _t _thr\ j_s p 5, 5.0].

If SNR2_lt_ave -_{g t}h_{an a set} threshold $NR2A, the spectral centroid feature ^sP_^cenl^erl^] is greater than a set threshold ^SP-^cenier_thrl ^d the long-time SNR It_snr is less than a set threshold //-/^-/-/^2, j_s determined that the current frame is an active frame and the flag vad flag is set to 1. Herein, a value range of SNR2_lt_a_Ve_t_thr3 _{is [0}.₈,2.0], a value range of ^p_cerUer _t_thr2 _{is [2}.₀,4.0], and a value range of is [2.5, 5.0].

If SNR2_lt_ave j_s gj._eat_er than a set threshold &amp;NR2 _lt _ave _t _thr4, _sp€Ctrai centroid feature ^SP _^center[2\ i_s greater than a set threshold sp _center _t _thr3 and the

CA 02990328 2017-12-20 long-time SNR It _snr is less than a set threshold _^tsnr.J_thr3, j_t j_{s de}t_ermi_ned that the current frame is an active frame and the flag vad_flag is set to 1. Herein, a value range of SNR2 _lt _ave _t _thrA _{is [0 6> 2 0]> a va|ue range of} sp_center_t_thr3 _{is [3 0> 6 0]? and}a value range of J™ _t _thr3 _{is[25> 5 0]}

In step 105f, the number of hangover frames for active sound is updated according to the decision results of several previous frames, the long-time SNR It _snr, and the average total SNR of all sub-bands SNR2_lt _ave and the VAD decision for the current frame.

Calculation steps are as follows.

The precondition for updating the current number of hangover frames for active sound is that the flag of active sound indicates that the current frame is active sound. If the condition is not satisfied, the current number of hangover frames num speech hangover is not updated and the process directly goes to step 105g.

Steps of updating the number of hangover frames are as follows.

If the number of continuous active frames continuous_speech_num is less than a set threshold continuous _speech_num_thr\ It Jsnr _{less than a}

It tsnr hthrl , the _current number of hangover frames for active sound num speech hangover _{is updated} by subtracting the number of continuous active frames continuous_speech_num from the minimum number of continuous active frames. Otherwise, if SNR2_lt _ave -_s g_{reater t}han _{a set} threshold SNR.2 Jt _ave _thr\ _{and the num}ber of continuous active frames continuous speech num is greater than a set second threshold continuous speech num_thr2 , _numb_{er o}f hangover frames for active sound num _speech Jhangover j_{s set accord}j_{ng t0} the value of *-^tsnr. Otherwise, the number of hangover frames num _speech ^hangover _is not updated. Herein, in the present embodiment, the minimum number of continuous active frames is 8, which may be between [6, 20], The first threshold eonlimous speech _md continuous _speech_num_thr2 _{may be same or different}

CA 02990328 2017-12-20

Steps are as follows.

If the long-time SNR It _snr is greater than 2.6, the value of num _ speech hangover is 3; otherwise, if the long-time SNR It _snr is greater than 1.6, the value of num _speech-hangover is 4; otherwise, the value of numspeech-hangover is 5.

In step 105g, a hangover of active sound is added according to the decision result and the number of hangover frames num _ speech hangover of the current frame, to obtain the VAD decision of the current frame.

The method thereof is as follows.

If the current frame is determined to be an inactive sound, that is the VAD flag is 0, and the number of hangover frames num_speech_hangover is greater than 0, the hangover of active sound is added, that is, the VAD flag is set to 1 and the value of num_speech_hangover is decreased by 1.

The final VAD decision of the current frame is obtained.

Alternatively, after step 105d, the following step may further be included : calculating the average energy of long-time active frames E_fg according to the initial VAD decision result, herein the calculated value is used for VAD decision of the next frame; and after step 105 g, the following step may further be included : calculating the average energy of long-time background noise E_bg according to the VAD decision result of the current frame, herein the calculated value is used for VAD decision of the next frame.

A calculation process of the average energy of long-time active frames E_fg is as follows;

a) if the initial VAD decision result indicates that the current frame is an active frame, that is, the value of the VAD flag is 1 and E_tl is many times (which is 6 times in the present embodiment) greater than E_bg, the cumulative value of average energy of long-time active frames fg_energy and the cumulative number of average energy of long-time active frames fg_energy_count are updated. An updating method is to add « to fg_energy to obtain a

CA 02990328 2017-12-20 new fg_energy, and add I to fg_energy_count to obtain a new fg_energy_count.

b) in order to ensure that the average energy of long-time active frames reflects the latest energy of active frames, if the cumulative number of average energy of long-time active frames is equal to a set value fg -frame * (he cumulative number and the cumulative value are multiplied by an attenuation coefficient ^attenu_^coeP _at the same time. In the present embodiment, a value of _^max_ frame num 512 and a value of attenu _ coef 1 · θ

c) the cumulative value of average energy of long-time active frames fg_energy is divided by the cumulative number of average energy of long-time active frames to obtain the average energy of long-time active frames, and a calculation equation is as follows:

£ _ fg_energy ^f&amp; fg_energy_count

A calculation method for the average energy of long-time background noise E^ is as follows.

It is assumed that bg_energy_count is the cumulative number of background noise frames, which is used to record how many frames of latest background noise are included in the energy cumulation. bg_energy is the cumulative energy of latest background noise frames.

a) if the current frame is determined to be a non-active frame, the value of the VAD flag is 0, and if SNR2 is less than 1.0, the cumulative energy of background noise bg_energy and the cumulative number of background noise frames bg_energy_count are updated. The updating method is to add the cumulative energy of background noise bgenergy to ¹¹ to obtain a new cumulative energy of background noise bg_energy. The cumulative number of background noise frames bg_energy_count is added by 1 to obtain a new cumulative number of background noise frames bg_energy_count.

b) if the cumulative number of background noise frames bg_energy_count is equal to the

CA 02990328 2017-12-20 maximum cumulative number of background noise frames, the cumulative number and the cumulative energy are multiplied by an attenuation coefficient attenucoe/2 at the same time. Herein, in this embodiment, the maximum cumulative number for calculating the average energy of long-time background noise is 512, and the attenuation coefficient attenu_coef2 is equal to 0.75.

c) the cumulative energy of background noise bg_energy is divided by the cumulative number of background noise frames to obtain the average energy of long-time background noise, and a calculation equation is as follows:

bgenergy bg_energy_count

In addition, it is to be illustrated that the embodiment one may further include the following steps.

In step 106, a background noise update flag is calculated according to the VAD decision result, the tonality feature, the SNR parameter, the tonality signal flag, and the time-domain stability feature of the current frame. A calculation method can be known with reference to the embodiment two described below.

In step 107, the background noise energy of the current frame is obtained according to the background noise update flag, the frame energy parameter of the current frame, and the energy of background noise over all sub-bands of the previous frame, and the background noise energy of the current frame is used to calculate the SNR parameter for the next frame.

Herein, whether to update the background noise is judged according to the background noise update flag, and if the background noise update flag is 1, the background noise is updated according to the estimated ratio of the energy of background noise over all sub-bands and the energy of the current frame. Estimation of the background noise energy includes both estimations of the sub-band background noise energy and estimation of energy of background noise over all sub-bands.

CA 02990328 2017-12-20

a. an estimation equation for the sub-band background noise energy is as follows:

^Esb2_bg W = ^Esb2_bg_pre (*) * ^abg_e ^{+ E}sb2_bg W * (¹ - ^ab_g_e W&lt;k&lt;HUm_sb

Herein, num _sb is the number of SNR sub-bands, and E_{sb2 bg} ^(k) represents the background noise energy of the k* SNR sub-band of the previous frame.

Λ ^bg-^e is a background noise update factor, and the value is determined by the energy of background noise over all sub-bands of the previous frame and the energy parameter of the current frame. A calculation process is as follows.

If the energy of background noise over all sub-bands E, _bg of the previous frame is less than the frame energy E_lt of the current frame, a value thereof is 0.96, otherwise the value is 0.95.

b. estimation of the energy of background noise over all sub-bands:

If the background noise update flag of the current frame is 1, the cumulative value of background noise energy E, and the cumulative number of background noise energy frames N_{Et cmnur} are updated, and a calculation equation is as follows:

E = F + F ·

N =N + 1· ' Et_coxmer ^J El _comter_-l ’

Herein, E, is the cumulative value of background noise energy of the previous frame, and N_El is the cumulative number of background noise energy frames calculated at the previous frame.

c. the energy of background noise over all sub-bands is obtained by a ratio of the cumulative value of background noise energy E, _sum and the cumulative number of frames N Ei_counter · jy /_sum '-^bs ~ ~N

Ei counter

CA 02990328 2017-12-20

It is judged whether N_{El counler} is equal to 64, if N_Et__couMer is equal to 64, the cumulative value of background noise energy E_t and the cumulative number of frames _Et__coumer are multiplied by 0.75 respectively.

d. the sub-band background noise energy and the cumulative value of background noise energy are adjusted according to the tonality signal flag, the frame energy parameter and the energy of background noise over all sub-bands. A calculation process is as follows.

If the tonality signal flag tonality _flag is equal to 1 and the value of the frame energy parameter £,, is less than the value of the background noise energy E_tbg multiplied by a gain coefficient gain,

E,^ = E, gain + delta·, E^^k) - E_sblbg(k)· gain + delta·,

Herein, a value range of gain is [0.3,1].

Embodiment two

The embodiment of the present invention further provides an embodiment of a method for detecting background noise, as shown in Fig. 3, including the following steps.

In step 201, a sub-band signal and a spectrum amplitude of a current frame are obtained.

In step 202, values of a frame energy parameter, a spectral centroid feature and a time-domain stability feature are calculated according to the sub-band signal, and values of a spectral flatness feature and a tonality feature are calculated according to the spectrum amplitude.

The frame energy parameter is a weighted cumulative value or a direct cumulative value of energy of all sub-band signals.

The spectral centroid feature is the ratio of the weighted sum to the non-weighted sum of energies of all sub-bands or partial sub-bands, or the value is obtained by performing smooth filtering on this ratio.

CA 02990328 2017-12-20

The time-domain stability feature is the ratio of the variance of the frame energy amplitudes to the expectation of the squared sum of energy amplitudes, or the ratio multiplied by a factor.

The spectral flatness feature is the ratio of the geometric mean to the arithmetic mean of predetermined certain spectrum amplitudes, or the ratio multiplied by a factor.

The same methods as above may be used in steps 201 and 202, and will not be described again.

In step 203, it is judged whether the current frame is background noise by performing background noise detection according to the spectral centroid feature, the time-domain stability feature, the spectral flatness feature, the tonality feature and the energy parameter of the current frame.

Firstly, it is assumed that the current frame is background noise and the background noise update flag is set to a first preset value; then, if any of the following conditions is met, it is determined that the current frame is not a background noise signal and the background noise update flag is set to a second preset value:

The time-domain stability feature lt_stable_rateO is greater than a set threshold.

The smoothed value of the spectral centroid feature is greater than a set threshold, and the time-domain stability feature is also greater than a set threshold.

A value of the tonality feature or a smoothed value of the tonality feature is greater than a set threshold, and a value of the time-domain stability feature lt_stable_rateO is greater than a set threshold.

A value of spectral flatness feature of each sub-band or smoothed value of spectral flatness feature of each sub-band is less than a respective corresponding set threshold,

Or, a value of the frame energy parameter E_tl is greater than a set threshold E_thr\.

Specifically, it is assumed that the current frame is a background noise.

CA 02990328 2017-12-20

In this embodiment, the background noise update flag background_flag is used to indicate whether the current frame is a background noise, and it is agreed that if the current frame is a background noise, the background noise update flag background_flag is set to 1 (the first preset value), otherwise, the background noise update flag background_flag is set to 0 (the second preset value).

It is detected whether the current frame is a noise signal according to the time-domain stability feature, the spectral centroid feature, the spectral flatness feature, the tonality feature, and the energy parameter of the current frame. If it is not a noise signal, the background noise update flag background_flag is set to 0.

The process is as follows.

It is judged whether the time-domain stability feature lt_stable_rateO is greater than a set threshold It_stable rate thrl. If so, it is determined that the current frame is not a noise signal and background flag is set to 0. In this embodiment, a value range of the threshold It_stable rate_thrl is [0.8,1.6];

It is judged whether the smoothed value of spectral centroid feature is greater than a set threshold sp center thrl and the value of time-domain stability feature is greater than a set threshold lt_stable_rate thr2. If so, it is determined that the current frame is not a noise signal and background_flag is set to 0. A value range of sp_center_thrl is [1.6, 4]; and a value range of lt_stable_rate_thr2 is (0,0.1].

It is judged whether the value of tonality feature (1) is greater than a set threshold tonality rate_thrl and the value of time-domain stability feature lt_stable_rateO is greater than a set threshold lt_stable_rate_thr3. If the above conditions are true at the same time, it is determined that the current frame is not a background noise, and the background flag is assigned a value of 0. A value range of the threshold _rate_thr\ is [0.4,0.66], and a value range of the threshold lt_stablc_rate_thr3 is [0.06,0.3 J.

It is judged whether the value of spectral flatness feature F^fO) is less than a set

CA 02990328 2017-12-20 threshold it is judged whether a value of the spectral flatness feature ^(1) is less than a set threshold and it is judged whether a value of the spectral flatness feature ^ (2) is less than a set value _ab_ove conditions are true at the same time, it is determined that the current frame is not a background noise, and the background_flag is assigned a value of 0, herein value ranges of the thresholds , sSMR_thr2 sSMRthrl β) 330 9g] j_t j_s judged whether the value of the spectral flatness feature ^(0) is less than a set threshold ._g whether the value of the spectral flatness feature /^(1) is less than a set threshold and it is judged whether the value of the spectral flatness feature /^(2) is less than a set threshold . if any of the above conditions is true, it is determined that the current frame is not a background noise. The background flag is assigned a value of 0. Value ranges of sSMRjhrt, sSMR_thr5 sSMR_thr6 _{are [0<80> 0}.92],

It is judged whether the value of frame energy parameter E_a is greater than a set threshold E_thr\. If the above condition is true, it is determined that the current frame is not F thr\ a background noise. The background_flag is assigned a value of 0. is assigned a value according to a dynamic range of the frame energy parameter.

If the current frame is not detected as non-background noise, it indicates that the current frame is a background noise.

Embodiment three

The embodiment of the present invention further provides a method for updating the number of hangover frames for active sound in VAD decision, as shown in Fig. 4, including the following steps.

In step 301, the long-time SNR It _snr is calculated according to sub-band signals.

The long-time SNR It _snr is obtained by computing the ratio of the average energy of long-time active frames to the average energy of long-time background noise for the previous

CA 02990328 2017-12-20 frame. The long-time SNR It _snr may be expressed logarithmically.

In step 302, the average total SNR of all sub-bands SNR2_lt_av^e j_s calculated.

The average total SNR of all sub-bands SNR2 _lt _ave\s obtained by calculating the average value of SNRs of all sub-bands SNR2 _s for multiple frames closest to the current frame.

In step 303, the number of hangover frames for active sound is updated according to the decision results, the long-time SNR It snr, and the average total SNR of all sub-bands SNR2_lt_ave, and the SNR parameters and the VAD decision for the current frame.

It can be understood that the precondition for updating the current number of hangover frames for active sound is that the flag of active sound indicates that the current frame is active sound.

For updating the number of hangover frames for active sound, if the number of continuous active frames is less than a set threshold 1 and the long-time SNR It_snr is less than a set threshold 2, the current number of hangover frames for active sound is updated by subtracting the number of continuous active frames from the minimum number of continuous active frames; otherwise, if the average total SNR of all sub-bands SNR2 _lt _ave is greater than a set threshold 3 and the number of continuous active frames is greater than a set threshold 4, the number of hangover frames for active sound is set according to the value of long-time SNR It _snr Otherwise, the number of hangover frames num _speech _hangover _{is no}, _updated.

Embodiment four

The embodiment of the present invention provides a method for acquiring the number of modified frames for active sound, as shown in Fig. 5, including the following steps.

In 401, a voice activity detection decision result of a current frame is obtained by the

CA 02990328 2017-12-20 method described in the embodiment one of the present invention.

In 402, the number of hangover frames for active sound is obtained by the method described in embodiment three of the present invention.

In 403, the number of background noise updates update_count is obtained. Steps are as follows.

In 403a, a background noise update flag background_flag is calculated with the method described in embodiment two of the present invention;

In 403b, if the background noise update flag indicates that it is a background noise and the number of background noise updates is less than 1000, the number of background noise updates is increased by 1. Herein, an initial value of the number of background noise updates is set to 0.

In 404, the number of modified frames for active sound warm_hang_num is acquired according to the VAD decision result of the current frame, the number of background noise updates, and the number of hangover frames for active sound.

Herein, when the VAD decision result of the current frame is an active frame and the number of background noise updates is less than a preset threshold, for example, 12, the number of modified frames for active sound is selected as the maximum number of a constant, for example, 20 and the number of hangover frames for active sound.

In addition, 405 may further be included: modifying the VAD decision result according to the VAD decision result, and the number of modified frames for active sound, herein:

When the VAD decision result indicates that the current frame is inactive and the number of modified frames for active sound is greater than 0, the current frame is modified as active frame and meanwhile the number of modified frames for active sound is decreased by 1.

CA 02990328 2017-12-20

Corresponding to the above method for acquiring the number of modified frames for active sound, the embodiment of the present invention further provides an apparatus 60 for acquiring the number of modified frames for active sound, as shown in Fig. 6, including the following units.

A first acquisition unit 61 is arranged to obtain the VAD decision of current frame.

A second acquisition unit 62 is arranged to obtain the number of hangover frames for active sound.

A third acquisition unit 63 is arranged to obtain the number of background noise updates.

A fourth acquisition unit 64 is arranged to acquire the number of modified frames for active sound according to the VAD decision result of the current frame, the number of background noise updates and the number of hangover frames for active sound.

The operating flow and the operating principle of each unit of the apparatus for acquiring the number of modified frames for active sound in the present embodiment can be known with reference to the description of the above method embodiments, and will not be repeated here.

Embodiment five

The embodiment of the present invention provides a method for voice activity detection, as shown in Fig. 7, including the following steps.

In 501, a first VAD decision result vada flag is obtained by the method described in the embodiment one of the present invention; and a second VAD decision result vadb flag is obtained.

It should be noted that the second VAD decision result vadb_flag is obtained with any of the existing VAD methods, which will not be described in detail herein.

In 502, the number of hangover frames for active sound is obtained by the method

CA 02990328 2017-12-20 described in the embodiment three of the present invention.

In 503, the number of background noise updates update_count is obtained. Steps are as follows.

In 503a, a background noise update flag background__flag is calculated with the method described in the embodiment two of the present invention.

In 503b, if the background noise update flag indicates that it is a background noise and the number of background noise updates is less than 1000, the number of background noise updates is increased by 1. Herein, an initial value of the number of background noise updates is set to 0.

In 504, the number of modified frames for active sound warm_hang_num is calculated according to the vada flag, the number of background noise updates, and the number of hangover frames for active sound.

Herein, when the vada_flag indicates an active frame and the number of background noise updates is less than 12, the number of modified frames for active sound is selected to be a maximum value of 20 and the number of hangover frames for active sound.

In 505, a VAD decision result is calculated according to the vadb_flag, and the number of modified frames for active sound, herein, when the vadb_flag indicates that the current frame is an inactive frame and the number of modified frames for active sound is greater than 0, the current frame is modified as an active frame and the number of modified frames for active sound is decreased by 1 at the same time.

Corresponding to the above VAD method, the embodiment of the present invention further provides a VAD apparatus 80, as shown in Fig. 8, including the following units.

A fifth acquisition unit 81 is arranged to obtain a first voice activity detection decision result.

CA 02990328 2017-12-20

A sixth acquisition unit 82 is arranged to obtain the number of hangover frames for active sound.

A seventh acquisition unit 83 is arranged to obtain the number of background noise updates.

A first calculation unit 84 is arranged to calculate the number of modified frames for active sound according to the first voice activity detection decision result, the number of background noise updates, and the number of hangover frames for active sound.

An eighth acquisition unit 85 is arranged to obtain a second voice activity detection decision result.

A second calculation unit 86 is arranged to calculate the VAD decision result according to the number of modified frames for active sound and the second VAD decision result

The operating flow and the operating principle of each unit of the VAD apparatus in the present embodiment can be known with reference to the description of the above method embodiments, and will not be repeated here.

Many modem voice coding standards, such as AMR, AMR-WB, support the VAD function. In terms of efficiency, the VAD of these encoders cannot achieve good performance under all typical background noises. Especially under an unstable noise, such as an office noise, the VAD of these encoders has low efficiency. For music signals, the VAD sometimes has error detection, resulting in significant quality degradation for the corresponding processing algorithm.

The solutions according to the embodiments of the present invention overcome the disadvantages of the existing VAD algorithms, and improve the detection efficiency of the VAD for the unstable noise while improving the detection accuracy of music. Thereby, better performance can be achieved for the voice and audio signal processing algorithms using the technical solutions according to the embodiments of the present invention.

In addition, the method for detecting a background noise according to the embodiment of

CA 02990328 2017-12-20 the present invention can enable the estimation of the background noise to be more accurate and stable, which facilitates to improve the detection accuracy of the VAD. The method for detecting a tonal signal according to the embodiment of the present invention improves the detection accuracy of the tonal music. Meanwhile, the method for modifying the number of hangover frames for active sound according to the embodiment of the present invention can enable the VAD algorithm to achieve better balance in terms of performance and efficiency under different noises and signal-to-noise ratios. At the same time, the method for adjusting a decision signal-to-noise ratio threshold in VAD decision according to the embodiment of the present invention can enable the VAD decision algorithm to achieve better accuracy under different signal-to-noise ratios, and further improve the efficiency in a case of ensuring the quality.

A person having ordinary skill in the art can understand that all or a part of steps in the above embodiments can be implemented by a computer program flow, which can be stored in a computer readable storage medium and is performed on a corresponding hardware platform (for example, a system, a device, an apparatus, and a component etc.), and when performed, includes one of steps of the method embodiment or a combination thereof.

Alternatively, all or a part of steps in the above embodiments can also be implemented by integrated circuits, can be respectively made into a plurality of integrated circuit modules; alternatively, it is implemented with making several modules or steps of them into a single integrated circuit module.

Each apparatus/fimctional module/functional unit in the aforementioned embodiments can be implemented with general computing apparatuses, and can be integrated in a single computing apparatus, or distributed onto a network consisting of a plurality of computing apparatuses.

When each apparatus/functional module/functional unit in the aforementioned embodiments is implemented in a form of software functional modules and is sold or used as

CA 02990328 2017-12-20 an independent product, it can be stored in a computer readable storage medium, which may be a read-only memory, a disk or a disc etc.

Industrial applicability

The technical solutions according to the embodiments of the present invention overcome the disadvantages of the existing VAD algorithms, and improve the detection efficiency of the VAD for the unstable noise while improving the detection accuracy of music. Thereby, the voice and audio signal processing algorithms using the technical solutions according to the embodiments of the present invention can achieve better performance. In addition, the method for detecting a background noise according to the embodiment of the present invention can enable the estimation of the background noise to be more accurate and stable, which facilitates to improve the detection accuracy of the VAD. Meanwhile the method for detecting a tonal signal according to the embodiment of the present invention improves the detection accuracy of the tonal music. At the same time, the method for modifying the number of hangover frames for active sound according to the embodiment of the present invention can enable the VAD algorithm to achieve better balance in terms of performance and efficiency under different noises and signal-to-noise ratios. The method for adjusting a decision signal-to-noise ratio threshold in VAD decision according to the embodiment of the present invention can enable the VAD decision algorithm to achieve better accuracy under different signal-to-noise ratios, and further improve the efficiency in a case of ensuring the quality.