KR20110127783A - Apparatus for separating voice and method for separating voice of single channel using the same - Google Patents

Abstract

PURPOSE: A voice separation apparatus and single channel voice separation method using the same are provided to improve the accuracy of a voice separation by considering the size of voice signal and phase information. CONSTITUTION: A voice signal division unit(110) separates phase information and size information from the mixed voice signal of a microphone. A size information probability unit(130) defines a soft mask filter. The size information probability unit calculates the DFT(Discrete Fourier Transform)] of the divided voice signal. A phase information probability unit(140) calculates a phase information probability. A voice signal extraction unit(150) estimates the desired voice signal.

Description

Apparatus for Separating Voice and Method for Separating Voice of Single Channel Using the Same}

The present invention relates to a voice separation device, and more particularly, to a voice separation device for separating a desired voice signal from a mixed voice signal in consideration of phase and magnitude information in a single channel, and a single channel voice separation method using the same.

Recently, with the development of digital signal processing technology, the related market is growing and various voice services such as voice dialing and locking device using voice communication and voice recognition system are becoming common.

The actual environment in which the speech recognition system is used includes several noise sources.

Therefore, techniques for mitigating the effects of noise are indispensable for improving speech signals and effectively applying them to speech information processing.

Voice separation technology, one of the approaches to remove noise, uses single-channel voice separation technology to perform voice separation using voice signals input from one microphone and two-channel voice separation using voice signals obtained from two microphones. Technology, and multi-channel speech separation using three or more microphones.

In a conventional speech processing system, a speech signal is obtained using one microphone.

Conventional single channel speech signal separation techniques include Computational Auditory Scene Analysis (CASA) and Soft Mask.

CASA is a method of separating a speech signal by finding acoustic elements generated from the same sound source by using human hearing characteristics from an input mixed speech signal.

Soft mask is a speech modeling method based on statistical modeling.

The voice separation method using CASA uses a binary mask as a sound source separation mask in the process of separating a sound source, resulting in a loss of a voice signal, which requires a phonetic knowledge and a heuristic.

In the speech separation method using a soft mask, the sound source separation mask uses a soft mask based on probability, but since the separation is performed by statistical modeling, there is a problem in that a non-contiguous case in which an audio signal is separated into other signals is generated even though it is an adjacent speech signal.

In addition, the soft mask has a problem in that the speech signal is separated using only the magnitude component (Magnitude) while ignoring the phase component having the information of the speech signal.

In order to solve this problem, the present invention separates the mixed speech signal input from the speech separation apparatus using magnitude information and phase information, and then extracts a desired speech signal from the mixed speech signal by combining the separated information. Its purpose is to.

Another object of the present invention is to compensate for the discontinuous signal separation process occurring in the speech separation process by applying a smoothing filter in the speech separation apparatus.

According to an aspect of the present invention, there is provided a single channel speech separation method comprising: separating magnitude and phase of the mixed speech signal to separate a desired target speech signal from a mixed speech signal input from a microphone; Calculating a magnitude probability that the magnitude of the mixed speech signal is the magnitude of the target speech signal using a magnitude model based on statistical modeling of magnitude information; Calculating a phase probability of the phase of the mixed speech signal being a phase of the target speech signal using a phase model based on statistical modeling of phase information; And estimating the target speech signal by multiplying the magnitude probability and the phase probability by the weight and the magnitude and phase of the mixed speech signal, respectively.

According to an aspect of the present invention, there is provided a voice separation device comprising: a voice signal separation unit for separating a magnitude and a phase of the mixed voice signal to separate a desired target voice signal from a mixed voice signal input from a microphone; A magnitude information probability section for calculating a magnitude probability that the magnitude of the mixed speech signal is the magnitude of the target speech signal using a magnitude model based on statistical modeling of magnitude information; A phase information probability unit for calculating a phase probability of the phase of the mixed speech signal being the phase of the target speech signal using a phase model based on statistical modeling of phase information; And a speech signal extractor configured to estimate the target speech signal by multiplying the magnitude probability and the phase probability by the weight and the magnitude of the mixed speech signal, respectively.

According to the above configuration, the present invention has an effect of increasing the accuracy of speech separation in consideration of both the magnitude and phase information of the speech signal during speech separation in the speech separation apparatus.

The present invention has an effect of compensating for the case of discontinuity caused by a statistical model by reflecting continuous features of speech using a smoothing filter.

The present invention has the effect of increasing the performance of speech separation by using the phase information and the size information in the speech processing field.

The present invention has the effect that can be utilized in the recognition of address names of telematics services in the home network environment, the noise reduction module of the voice interface of the intelligent robot.

FIG. 1 is a block diagram schematically illustrating an internal configuration of a speech separation apparatus for separating a desired speech signal using a magnitude and phase of speech in a single channel according to an embodiment of the present invention.
2 is a schematic diagram showing a single channel speech separation method in a speech separation apparatus according to an embodiment of the present invention.
3 is a diagram showing an example of outputting log spectra and phase components for a part of one frame of an input speech signal x (t), y (t) and a mixed speech signal z (t) according to an embodiment of the present invention.

DETAILED DESCRIPTION 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. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and like reference numerals designate like parts throughout the specification.

Throughout the specification, when a part is said to "include" a certain component, it means that it can further include other components, without excluding other components unless specifically stated otherwise.

An embodiment of the present invention separates a speech signal in consideration of phase and magnitude information in a single channel.

FIG. 1 is a block diagram schematically illustrating an internal configuration of a speech separation apparatus for separating a desired speech signal using a magnitude and phase of speech in a single channel according to an embodiment of the present invention.

In the speech separation apparatus 100 according to the present exemplary embodiment, the speech signal separation unit 110, the smoothing filter unit 120, the magnitude information probability unit 130, the phase information probability unit 140, and the speech signal extraction unit 150 are described. It includes.

The voice signal separating unit 110 separates magnitude information and phase information from the mixed voice signal input from the microphone.

In the present invention, a magnitude model for separating a desired speech signal from a mixed speech signal using magnitude information is calculated using a soft mask, which is one of speech separation methods based on conventional statistical modeling.

The soft mask calculates a probability that the input mixed signal is a desired signal, and then multiplies the calculated probability value by the mixed signal again to estimate a desired speech signal.

의 입력 음성 신호를 각각 x(t), y(t)라고 할 때, 혼합된 음성 신호 z(t)는 두 입력 음성 신호의 합으로 다음의 [수학식 1]과 같다.Speaker obtained through a single microphone

When the input speech signals of are x (t) and y (t), respectively, the mixed speech signal z (t) is the sum of the two input speech signals as shown in Equation 1 below.

Here, it is assumed that x (t) and y (t) are signals independent of each other. If the log spectra of x (t) and y (t) are x (w) and y (w), the log spectrum z (w) of the mixed speech signal is calculated as x (w) + y (w). It is redefined as in [Equation 2].

In general, the log spectrum of the mixed speech signal is very similar to the log spectrum having the larger of the log spectra of the two input speech signals.

Therefore, the log spectrum of the mixed speech signal can be approximated using Equation 3 using Log-Max Approximation.

According to the logarithmic approximation defined in Equation 3, the probability that the log spectrum vector of the mixed speech signal is the desired speech signal x is equal to the probability that the value of the log spectrum of x is greater than the value of y. That is, the log spectrum of the d-th order of the value z z _d x _d is the probability is calculated as the probability that the value of x is greater than y _{d d} shown in the following [Equation 4].

The logarithm of the desired signal is extracted by applying a probability value calculated using Equation 4 as a weight to the logarithm of the mixed signal.

Since the result of estimating the desired speech signal using the soft mask is obtained from the assumption that the two speakers x and y are independent signals, the log spectrum z of the mixed speech signal is separated into either x or y.

Thus, log spectral separation results in the immediately next time-frequency band can be calculated with different speech signals, resulting in discontinuous cases.

To compensate for this, smoothing is applied to the speech signal using a smoothing filter, and then the probability is calculated to increase the continuity between adjacent time-frequency bands and to consider the characteristics of the speech.

The smoothing filter unit 120 smoothes the current frame (d th frame) using neighboring log spectral values in the time-frequency domain when calculating a probability that the log spectral value z _d of the mixed speech signal is x _d .

이다.The mask for calculating smoothing uses a uniform mask filter (△ = 4) with a uniform filter width in the time-frequency domain, and the size of the filter width is

to be.

The smoothing filter unit 120 calculates the log spectrum vector to which smoothing is applied through Equation 5 below.

은 다음의 [수학식 5]와 같다.If the log spectral vector with smoothing in the time-frequency domain is z ', the k-th log spectral value of z'

Is shown in Equation 5 below.

,

라 하면 전술한 수학식 5와 같은 과정을 거쳐 계산한다.Log spectrum with smoothing in time-frequency domain to log spectra of two speech signals x and y

,

If it is calculated through the same process as the above equation (5).

라 하면, 스무딩이 적용된 소프트 마스트 필터를 다음의 [수학식 6]과 같이 정의한다.The magnitude information probability unit 130 calculates a log spectral vector to which smoothing is applied.

In this case, the soft mast filter to which smoothing is applied is defined as in Equation 6 below.

)은 전술한 [수학식 4]와 베이시안 정리(Bayesian Theory)를 이용하여 계산한다.Probability that the mixed speech signal is the desired speech signal (

) Is calculated using the above Equation 4 and the Bayesian Theory.

Here, m _x and m _y are indices of weighted sums corresponding to the number of Gaussians M _x and M _y in the mixed Gaussian model of _x and _y .

과 분산

가 주어졌을때 가우시안 분포를

라 하면, x'의 m_x번째 가우시안의 평균 벡터의 d번째 평균값

와 분산 벡터의 d번째 분산값

가 주어졌을때 확률 밀도 함수

는 다음의 [수학식 7]과 같고, y'의 m_y번째 가우시안의 평균 벡터의 d번째 평균값

와 분산 벡터의 d번째 분산값

가 주어졌을때의 누적 확률 밀도

는 다음의 [수학식 8]과 같다.Parameters such as mean, variance, etc. are calculated from the learning speech data

And dispersion

Given the Gaussian distribution

When La, x 'of the m _x-th mean vector of the Gaussian second average value of d

D variances of and the variance vector

Given the probability density function

Is equal to Equation 7 below, and the d-th average of the mean vector of the _y- th Gaussian of y '

D variances of and the variance vector

Cumulative probability density given by

Is shown in Equation 8 below.

라고 하면, 혼합 음성 신호가 원하는 신호일 확률의 d번째 확률값

은 다음의 [수학식 9]와 같이 정의된다.The probability value calculated using the above Equation 6

Is the d-th probability value of the probability that the mixed speech signal is a desired signal.

Is defined as in Equation 9 below.

는 확률값

와 혼합 음성 신호의 로그 스펙트럼 벡터의 d번째 로그 스펙트럼 z_d의 곱으로 계산되며 다음의 [수학식 10]과 같다.D-d log spectrum of log spectral vector estimated by speaker S _x from mixed speech signal

Is a probability value

And is calculated as the product of the d th log spectrum z _d of the log spectral vector of the mixed speech signal, as shown in Equation 10 below.

)을 로그 스펙트럼 벡터

와 혼합 음성 신호의 위상 스펙트럼 벡터를 이용하여 계산한다.The magnitude information probability unit 130 performs a discrete Fourier transform (DFT) of the separated speech signal (

Log spectrum vector

And using the phase spectrum vector of the mixed speech signal.

일 때 다음의 [수학식 11]과 같이 계산된다.The phase component of the mixed speech signal z (t)

Is calculated as in Equation 11 below.

In the present invention, the desired speech signal is estimated by calculating the probability that the phase information vector w of the mixed speech signal z (t) is the phase information of the desired speech signal and multiplying the mixed signal.

In the present invention, assuming that the phase spectrum vectors of the two voice signals x (t) and y (t) are w, it is assumed that the phase of the mixed voice signal is similar to the phase of one of the two signals. 12].

)을 위상 스펙트럼값들 사이의 차이 정보를 이용하여 계산한다(아래의 [수학식 13]에서 정의됨).The phase information probability unit 140 may determine the probability that the phase spectrum vector w _d of the mixed speech signal is the phase u _d of the desired speech signal x (

) Is calculated using the difference information between phase spectral values (defined in Equation 13 below).

).The probability that the phase spectral vector of the mixed speech signal is the phase of the desired speech signal is equal to the probability that the similarity value between the phase spectrum u _d of w _d and x is higher than the similarity value between the phase spectrum v _d of w _d and y (

).

After extending Equation 13 using Bayesian theorem, the range is calculated to remove the absolute value operator, as shown in Equation 14 below.

)은 다음의 [수학식 15]와 같다.From Equation 14, the probability that wd is the phase spectrum u _d of the desired signal (

) Is as shown in Equation 15 below.

을

라고 하면, d번째 확률값

와 혼합 음성 신호의 위상 정보로부터 추정된 로그 스펙트럼 벡터

와 위상 스펙트럼 벡터

의 d번째 값

는 다음의 [수학식 16], [수학식 17], [수학식 18]과 같다.Probability that the phase spectrum of the mixed speech signal obtained using the phase model is the phase spectrum of the speaker S _x

of

If we say d, the probability value

Spectral vector estimated from the phase information of the mixed speech signal

And phase spectrum vector

D value of

Is the same as [Equation 16], [Equation 17], [Equation 18].

)을 다음의 [수학식 19]를 통해 계산한다.The phase information probability unit 140 performs discrete Fourier transform of the speech signal estimated using the log spectrum vector and the phase spectrum vector.

) Is calculated by the following Equation 19.

In the process of separating a desired speech signal from the mixed speech signal, the speech signal extractor 150 uses the magnitude model and the phase model to calculate the speech information calculated using the magnitude model and the phase model. The desired speech signal is estimated by multiplying the information and the phase information by weight.

, 위상 모델로부터 얻어진 음성 신호의 이산 푸리에 변환을

라고 하면, 음성 신호의 크기 정보와 위상 정보를 고려하여 분리된 음성 신호는 다음의 [수학식 20]와 같이 정의된다.Discrete Fourier Transform of Speech Signal Extracted from Magnitude Model

Discrete Fourier Transform of Speech Signal Obtained from Phase Model

In this regard, the separated speech signal in consideration of the magnitude information and the phase information of the speech signal is defined as in Equation 20 below.

)를 역변환한 후, 오버랩-애드 방법을 사용하여 음성 신호를 복원하여 원하는 음성 신호를 얻는다.The voice signal extractor 150 is a separated voice signal (

) Is then inverted and the speech signal is recovered using the overlap-add method to obtain the desired speech signal.

Referring to FIG. 2, the single channel voice separation method in the voice separation device 100 will be described in detail.

2 is a schematic diagram showing a single channel speech separation method in the speech separation apparatus 100 according to an embodiment of the present invention.

2 illustrates a speech separation method of estimating a desired speech signal by combining the estimated speech signal using the magnitude and phase information of the speech signal.

The voice signal separating unit 110 separates magnitude information and phase information from the mixed voice signal input from the microphone.

The smoothing filter unit 120 smoothes the current frame (d th frame) using neighboring log spectral values in the time-frequency domain when calculating a probability that the log spectral value z _d of the mixed speech signal is x _d .

The magnitude information probability unit 130 calculates a probability that the magnitude of the mixed speech signal is the magnitude of the desired signal in order to estimate the desired speech signal using the magnitude model in the mixed speech signal. The method for calculating the probability of magnitude has been described with reference to FIG. 1.

Here, the size model is a method of applying a mask based on probability using a soft mask based on statistical modeling.

 The phase information probability unit 140 calculates a probability that the mixed speech signal is a phase of the desired speech signal in order to estimate the desired speech signal using the phase model in the mixed speech signal. The method of calculating the probability of being a phase has been described with reference to FIG. 1.

Here, the phase model is a method using a soft mask applied to the magnitude model for the phase component such that the magnitude component of the mixed speech signal exhibits a phase similar to the phase of the largest signal.

The speech signal extractor 150 multiplies the probability calculated by the magnitude information probability unit 130 and the probability calculated by the phase information probability unit 140 by multiplying the magnitude and phase of the mixed speech signal by weight, respectively, to combine the desired speech signal. Extract.

An example of the output of the log spectrum and phase component for a part of one frame of the input speech signal x (t), y (t) and the mixed speech signal z (t) will be described with reference to FIG.

3 is a diagram showing an example of outputting log spectra and phase components for a part of one frame of an input speech signal x (t), y (t) and a mixed speech signal z (t) according to an embodiment of the present invention.

As shown in FIG. 3, in the period x where the log spectrum of x (t) is larger than the log spectrum of y (t), the phase spectrum of the mixed speech signal appears similar to the phase spectrum of x (t).

In other words, just as the log spectrum, which is the magnitude component of the mixed speech signal, is similar to the larger of the two input speech signals, the phase component of the mixed speech signal exhibits a value very similar to the phase of the signal having the larger magnitude component.

In the interval y where the log spectrum of x (t) is smaller than the log spectrum of y (t), the phase spectrum of the mixed speech signal appears similar to the phase spectrum of y (t).

The embodiments of the present invention described above are not implemented only by the apparatus and / or method, but may be implemented through a program for realizing functions corresponding to the configuration of the embodiment of the present invention, a recording medium on which the program is recorded And such an embodiment can be easily implemented by those skilled in the art from the description of the embodiments described above.

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements of those skilled in the art using the basic concepts of the present invention defined in the following claims are also provided. It belongs to the scope of rights.

Claims (11)

Separating the magnitude and phase of the mixed speech signal to separate a desired target speech signal from the mixed speech signal input from the microphone;
Calculating a magnitude probability that the magnitude of the mixed speech signal is the magnitude of the target speech signal using a magnitude model based on statistical modeling of magnitude information;
Calculating a phase probability of the phase of the mixed speech signal being a phase of the target speech signal using a phase model based on statistical modeling of phase information; And
Estimating the target speech signal by multiplying the magnitude probability and the phase probability by the weight and the magnitude and phase of the mixed speech signal, respectively.
Single channel voice separation method comprising a. The method of claim 1,
The step of calculating the size probability,
Applying the smoothing using the log spectral value of the neighboring frame to the current frame and calculating the magnitude probability based on the smoothed log spectral value.
Single channel voice separation method further comprising. The method of claim 1,
The step of calculating the size probability,
Calculating a second log spectral vector of the target speech signal by multiplying the first log spectral vector of the mixed speech signal by the magnitude probability; And
Calculating a first discrete Fourier transform of the speech signal extracted from the magnitude model using the second log spectrum vector and the phase spectrum vector of the mixed speech signal
Single channel voice separation method comprising a. The method of claim 3,
Computing the phase probability,
Calculating a third log spectral vector of the target speech signal by multiplying the first log spectral vector of the mixed speech signal by the phase probability;
Calculating a phase spectral vector of the target speech signal by multiplying the phase spectrum vector of the mixed speech signal by the phase probability; And
Calculating a second discrete Fourier transform of the speech signal extracted from the phase model using the third log spectrum vector and the phase spectrum vector of the target speech signal
Single channel voice separation method comprising a. The method of claim 4, wherein
Estimating the target speech signal,
Estimating the target speech signal by combining the magnitude probability, the phase probability, the first discrete Fourier transform, and the second discrete Fourier transform;
Single channel voice separation method comprising a. The method of claim 1,
Wherein the magnitude probability is calculated using a probability that a value of a log spectrum of the target speech signal is greater than a value of a log spectrum of a speech signal other than the target speech signal. The method of claim 1,
The phase probability is that the similarity value between the phase spectrum vector of the mixed speech signal and the phase spectrum vector of the target speech signal is similar to the phase spectrum vector of the speech signal other than the target speech signal. Single channel speech separation method, characterized in that the probability higher than the value. A voice signal separation unit for separating the magnitude and phase of the mixed voice signal to separate a desired target voice signal from the mixed voice signal input from the microphone;
A magnitude information probability section for calculating a magnitude probability that the magnitude of the mixed speech signal is the magnitude of the target speech signal using a magnitude model based on statistical modeling of magnitude information;
A phase information probability unit for calculating a phase probability of the phase of the mixed speech signal being the phase of the target speech signal using a phase model based on statistical modeling of phase information; And
A speech signal extractor for estimating the target speech signal by multiplying the magnitude probability and the phase probability by the weight and the magnitude and phase of the mixed speech signal, respectively.
Voice separation apparatus comprising a. The method of claim 8,
The smoothing filter unit applies smoothing using the log spectral vectors of neighboring frames using the smoothing filter and transmits the smoothed log spectral vector as an input signal of the magnitude information probability unit for calculating the magnitude probability.
Voice separation device further comprising. The method of claim 8,
And the magnitude information probability unit calculates the magnitude probability using a probability that a value of a log spectrum of the target speech signal is greater than a value of a log spectrum of an unwanted speech signal among the mixed speech signals. The method of claim 8,
The magnitude model is a method of applying a mask based on probability using a soft mask based on statistical modeling, and the phase model is a phase whose value is similar to that of a signal having the largest magnitude component of the mixed speech signal. And a soft mask applied to the magnitude model to represent the phase component so as to represent.

Images