KR101124712B1 - A voice activity detection method based on non-negative matrix factorization - Google Patents

Abstract

The present invention relates to a speech detection method based on non-negative matrix factorization, and more particularly, to a speech detection method, comprising: (1) deriving a basis vector from a noise signal and an input signal using a non-negative matrix factorization technique; (2) calculating an error between the noise signal base vector and the input signal base vector derived in step (1), and (3) applying a threshold to the error calculated in step (2) And including the step of detecting.
According to the speech detection method based on the non-negative matrix factorization proposed in the present invention, the error between the basis vectors extracted from the noise signal and the input signal is calculated, and the active section of the speech is classified based on the same. The noise threshold is estimated according to the distribution of error values, and the optimal threshold value is selected and applied to the error to detect the speech signal. Therefore, the signal-to-noise ratio is low, so that the speech detection can be performed in a relatively poor noise environment.

Description

Speech detection method based on non-negative matrix factorization {A VOICE ACTIVITY DETECTION METHOD BASED ON NON-NEGATIVE MATRIX FACTORIZATION}

The present invention relates to a speech detection method, and more particularly, to a speech detection method based on non-negative matrix factorization.

Voice activity detectors (VADs), which detect voice and non-voice intervals, are widely used in voice communication systems such as voice encoding, speech recognition, and acoustic echo canceller. In particular, it is essential to use the bandwidth of voice communication system efficiently and statistical model for the existence and absence of speech used in the minimum mean square error (MMSE) based speech enhancement technique, which was recently started in Ephraim and Malah's research, It is known that the application to the voice detector has excellent performance.

Conventional speech detection algorithms determine whether speech is active based on statistical models of speech and noise, so that it is possible to detect speech more accurately in signals with high signal-to-noise ratios, but the performance of speech detection is rapidly reduced in relatively poor noise environments. There is a disadvantage of deterioration.

On the other hand, non-negative matrix factorization is an algorithm that shows excellent performance in pattern learning and pattern recognition of images. In particular, it is useful for data feature extraction because the optimal basic pattern can be separated from a plurality of input data and the entire data can be approximated by a linear combination thereof. However, an efficient speech detection method using non-negative matrix factoring has not been proposed yet.

The present invention has been proposed to solve the above problems of the conventionally proposed methods, and calculates an error between an input signal and a basis vector extracted from a noise signal and classifies an active section of a speech based on the noise estimation. Non-negative matrix, which estimates the noise environment according to the error value distribution in the interval, selects the optimal threshold value, and applies the error to detect the speech signal. It is an object of the present invention to provide a factor detection-based speech detection method.

In order to achieve the above object, a non-negative matrix factorization-based speech detection method according to a feature of the present invention,

(1) deriving a basis vector from the noise signal and the input signal using a non-negative matrix factorization technique;

(2) calculating an error between the noise signal base vector and the input signal base vector derived in step (1); And

(3) detecting the speech by applying a threshold value to the error calculated in the step (2).

Preferably,

In the absence of speech estimated by the statistical model-based speech detection method, an optimized threshold value is obtained according to the distribution of error values of the noise signal and the input signal derived by using the non-negative matrix factorization method.

The voice may be detected using the threshold value of step (3) as the optimized threshold value.

More preferably, the optimized threshold value is

Each noise environment may have a different value.

Preferably,

The basis vector may be a posteriori signal-to-noise ratio (SNR) and a priori SNR estimated using a statistical model.

According to the speech detection method based on the non-negative matrix factorization proposed in the present invention, the error between the basis vectors extracted from the noise signal and the input signal is calculated, and the active section of the speech is classified based on the same. The noise threshold is estimated according to the distribution of error values, and the optimal threshold value is selected and applied to the error to detect the voice signal. Thus, the low-to-noise ratio enables excellent voice detection even in a relatively poor noise environment.

1 is a flow diagram illustrating a non-negative matrix factorization-based speech detection method according to an embodiment of the present invention.
2 is a diagram illustrating a distribution of non-negative matrix factorization results in a section estimated as a noise environment through a statistical model-based speech detection method.
FIG. 3 is a diagram illustrating non-negative matrix factorization based speech detection results for noise to speech signal ratios in a white noise environment. FIG.
4 is a diagram illustrating a change in voice detection probability according to a noise environment.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. However, in describing the preferred embodiment of the present invention in detail, if it is determined that the detailed description of the related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. In addition, the same reference numerals are used throughout the drawings for parts having similar functions and functions.

In addition, in the entire specification, when a part is referred to as being 'connected' to another part, it may be referred to as 'indirectly connected' not only with 'directly connected' . In addition, the term 'comprising' of an element means that the element may further include other elements, not to exclude other elements unless specifically stated otherwise.

1 is a diagram illustrating a flow of a voice detection method based on non-negative matrix factorization according to an embodiment of the present invention. As shown in FIG. 1, in the non-negative matrix factorization-based speech detection method according to an embodiment of the present invention, a step of deriving a basis vector from a noise signal and an input signal (S100), the noise signal base vector, and an input is performed. The method may include the step of calculating an error between the signal base vectors (S200), and the step of detecting a voice (S400), and may further include the step of obtaining an optimized threshold value (S300).

That is, in the speech detection method based on non-negative matrix factorization according to an embodiment of the present invention, first, a base vector is derived from a noise signal and an input signal using a non-negative matrix factorization technique (S100). An error between the derived noise signal base vector and the input signal base vector is calculated (S200), and a voice is detected by applying a threshold value to the calculated error (S400). In this case, an optimized threshold value may be used to improve speech detection performance. To this end, a noise signal and an input obtained by using a non-negative matrix factorization method in a speech absence section estimated by a statistical model-based speech detection method may be used. An optimized threshold value according to an error value distribution of the signal may be obtained (S300).

Hereinafter, the details of the non-negative matrix factorization-based speech detection method according to an embodiment of the present invention will be described in detail.

For the speech detection method based on non-negative matrix factorization according to an embodiment of the present invention, analysis of the speech detection method based on a statistical model should be preceded. First, in the speech detection method based on a statistical model, the input signal y (t) to which the noise signal n (t) is applied to the original speech signal x (t) on the time axis is frequencyd through a discrete Fourier transform (DFT). Convert it to an axis and express it as in Equation 1 below.

Where Y (t) = [Y ₁ , Y ₂ ,... , Y _M ], X (t) = [X ₁ , X ₂ ,... , X _M ] and N (t) = [N ₁ , N ₂ ,... , N _M ] represent the DFT coefficient vectors of the speech signal contaminated with noise, the original speech signal, and the noise signal, respectively. Given that the hypotheses H ₀ and H ₁ represent the absence and presence of speech, the following equations are described for each frequency channel.

From the assumption that the spectrums of speech and noise signals follow a complex Gaussian distribution, the probability density function under the hypotheses H ₀ and H ₁ is given by Equation 3 below.

Here, λ _{x, k} and λ _{d, k} are the variances of speech and noise for each channel, and the likelihood ratio for the k th frequency band is obtained as in Equation 4 below.

Here, ξ _k = λ _{x, k} / λ _{d, k} and γ _k = Y _k / λ _{d, k} are a priori signal-to-noise ratio (SNR) and a posteriori SNR, respectively. A posteriori SNR γ _k is estimated using the noise variance λ _{d, k} obtained from the noise signal updated in the speech-free interval, and a priori SNR ξ _k is determined using the decision-directed (DD) method Estimate together.

은 이전 프레임에서 추정된 음성 신호의 k번째 스펙트럼 성분의 크기에 대한 추정치이며, MMSE에 기반을 두어 구한다. 또한 α는 가중치 값이며, 연산자 P[?]은 다음 수학식 6과 같이 정의된다.here,

Is an estimate of the magnitude of the k th spectral component of the speech signal estimated in the previous frame and is obtained based on the MMSE. Α is a weight value, and the operator P [?] Is defined as in Equation 6 below.

In the conventional general statistical model-based speech detector, a geometric mean of the likelihood ratios obtained in each frequency channel is used to determine whether speech is detected, as shown in Equation 7 below.

Where M is the number of all frequency bands, and η is the voice detection threshold.

In the present invention, a posteriori SNR estimated using the noise variance λ _{d, k} as described above and a priori estimated using the DD scheme SNR can be used as a base vector.

In step S100, a base vector is derived from the noise signal and the input signal using a non-negative matrix factorization technique.

형태, 즉 다음 수학식 8과 같이 인수분해된다.Non-negative matrix factorization (NMF), like PCA and VQ, is a technique that decomposes a matrix by approximating a linear combination of basis vectors. NMF has a restriction that all components are non-negative. The n × m matrix V of nonnegative components

Form, that is, factored as

Here, the matrix V can be regarded as m of n x 1 vectors combined, the matrix W is a set of r n x 1 vectors of size n × r, and H is a coefficient matrix of size r × m. That is, m n-th data vectors are expressed as a linear combination of r n-th basis vectors.

이며 이것을 최소화하는 성분 값 갱신 식은 다음 수학식 9와 같다.In general, the nonnegative matrix factorization process of the matrix is approximated in a direction to minimize the distance between the matrix V and WH by repeatedly updating the matrix W and H. Therefore, the component value update expression varies depending on which distance function or objective function is applied. Commonly used objective function is Euclidean distance.

The component value update equation for minimizing this is as shown in Equation 9.

In the present invention, a posteriori SNR estimated using the noise variance λ _{d, k} as described above and a priori estimated using the DD scheme With the SNR as the basis vector, assuming that m + 9 frames at the beginning are speech absent intervals, the superframes of m frames are non-negative matrix factorized to extract the base vector W of the background noise signal, and then, one frame at a time. The base vector can be extracted by moving.

In step S200, an error between the noise signal base vector and the input signal base vector derived in step S100 is calculated. Non-negative matrix factorization approximates a data vector as a linear combination of base vectors. Therefore, the similarity between the input signal and the noise signal may be determined based on a value obtained by extracting a base vector from the noise signal in advance and calculating a distance from the base vector of the input signal. Accordingly, by calculating the error between the noise signal base vector and the input signal base vector, the similarity according to the error magnitude is determined to detect the voice in the input signal.

In operation S300, an optimized threshold value according to a distribution of error values of a noise signal and an input signal derived by using a non-negative matrix factorization method in a speech member section estimated by a statistical model based speech detection method is obtained.

In the present invention, a statistical model-based speech detector is introduced to apply a threshold value optimized for various noise environments. In order to recognize the noise environment using the NMF result and apply the optimal threshold value corresponding to the noise environment, it is desirable to exclude the speech section as much as possible. In order to effectively estimate the noise section, we apply a statistical model-based speech detector to the starting N frame and classify the noise environment by the average of NMF results in case of non-voice.

FIG. 2 is a diagram illustrating a distribution of non-negative matrix factorization results in a section estimated as a noise environment through a statistical model-based speech detection method. As shown in FIG. 2, the NMF result distribution can be clearly distinguished according to each noise environment. Therefore, the threshold value optimized for each noise environment is applied as in Equation 10 below.

Where N is the frame size for estimating the noise environment, and b is the number of frames in the N frame where logΛ (t) < It should be noted that statistical model-based speech detection algorithm can be used to apply the optimized threshold to the noise environment to obtain better speech detection results. It can be applied in a way.

In step S400, a voice is detected by applying a threshold to the error calculated in step S200. In this case, using the optimized threshold value η derived in step S300 can improve the performance of the voice detection method according to an embodiment of the present invention.

와 입력 신호의 기초 벡터들 사이의 거리를 계산하여 음성 활성 여부를 다음과 같이 판단한다. 특히, 단계 S300에서 도출된 최적화된 문턱 값 η를 다음 수학식 12에 적용하여 최종적으로 음성을 검출할 수 있다.Total of 10 basis vectors

The distance between the basis vectors of the input signal is calculated to determine whether the voice is active as follows. In particular, the optimized threshold value η derived in step S300 may be applied to Equation 12 to finally detect speech.

는 배경 잡음 신호로부터 추출된 것이므로 이 기초 벡터들과 유사한 신호가 입력될 경우 거리는 상대적으로 작으며, 잡음 신호와 특성이 다른 음성 신호와의 거리는 상대적으로 매우 크다.
Extracted basis vector

Since is extracted from the background noise signal, when a signal similar to the basis vectors is input, the distance is relatively small, and the distance between the noise signal and the speech signal having different characteristics is relatively large.

와 입력 신호의 기초벡터 W사이의 거리를 보여주고 있다.
FIG. 3 is a diagram illustrating a result of speech detection based on non-negative matrix factor of noise to speech signal ratio in a white noise environment. 3 shows the results when the SNR is 5dB of white noise. 3 is a basis vector of a noise signal in a white noise environment

And the distance between the base vector W and the input signal.

는 입력신호가 동일한 잡음 환경이라면 0에 가까운 값을 가져야 하지만 그렇지 않다. 이것은 잡음이 stationary 혹은 non-stationary의 정도에 따라 잡음 신호의 기초 벡터와 입력 신호의 기초 벡터 간의 차이가 다른 값을 가지기 때문이다.
4 is a diagram illustrating a change in voice detection probability according to a noise environment. 4 shows a variation curve of the speech detection probability P _e (speech detection error probability) according to the change of the threshold value η 'in the NMF-based speech detector. As shown in FIG. 2, it can be seen that the NMF results have respective optimized threshold values according to the noise environment. Therefore, it is desirable to have a threshold value according to the noise environment for optimized performance. Base Vector in Noise Environment Extracted from Starting Frame Assumed as Speech Absence Section

Must be close to zero if the input signal is in the same noise environment, but it is not. This is because the difference between the basis vector of the noise signal and the basis vector of the input signal is different depending on the degree of noise stationary or non-stationary.

Experiment result

In order to evaluate the performance of the proposed voice detection algorithm, the voice detection error probability P _e including false alarm and missing was measured. The data used in the experiment was to manually display the voice and non-voice parts every 10 ms in clear voice data of 230 seconds in total. The voice section of the classified voice data was composed of 57.1% of total voice, 44.0% of voiced sound and 13.1% of unvoiced sound. , 5, 10, 15 kHz SNR was used in addition to the original audio signal. The feature vectors used to find the basis vector in NMF were 32 orders of a priori SNR ξ _k , a posteriori SNR γ _k , and m = 5 and r = 3. The noise basis vector was used as the average of a total of 10 basis vectors moving by one frame, and the threshold value was determined by dividing the noise environment by the mean value of the distribution using the NMF result distribution in 50 frames to estimate each noise. .

The results of the negative detection experiments are shown in Table 1 below. Experimental results show that the low threshold SNR results in a significant performance improvement for the NMF-based speech detector. It can be seen that the NMF-based speech recognizer is less affected by the signal-to-noise ratio than the statistical speech-based speech recognizer.

The present invention described above may be variously modified or applied by those skilled in the art, and the scope of the technical idea according to the present invention should be defined by the following claims.

S100: deriving a basis vector from the noise signal and the input signal
S200: calculating an error between the noise signal base vector and the input signal base vector
S300: calculating the optimized threshold value
S400: detecting voice

Claims (4)

As a voice detection method,
(1) deriving a basis vector from the noise signal and the input signal using a non-negative matrix factorization technique;
(2) calculating an error between the noise signal base vector and the input signal base vector derived in step (1); And
(3) detecting a voice by applying a threshold to the error calculated in step (2),
In the speech component section estimated by the statistical model-based speech detection method, the optimized threshold value is calculated according to the distribution of error values of the noise signal and the input signal derived by using the non-negative matrix factorization method.
Non-negative matrix factorization-based speech detection method characterized in that for detecting the speech using the threshold value of the step (3) as the optimized threshold value.
delete The method of claim 1, wherein the optimized threshold value,
Non-negative matrix factorization-based speech detection method characterized in that different values for each noise environment.
The method of claim 1,
The basis vector is a negative posterior signal-to-noise ratio (SNR) and a priori SNR estimated using a statistical model.

Images