KR101082837B1 - Method and apparatus for reduction of noise - Google Patents

Abstract

The present invention relates to an apparatus and a method for removing noise. In order to improve the noise removal efficiency in an environment in which noise fluctuations are mixed and various kinds of noise, voice and noise separation techniques such as a soft masking technique, etc. By using the noise adaptation technique to enhance the separation function of the noise and to compensate for the limitation that the noise Gaussian mixture model has to model the noise component of the input signal, it is possible to estimate the clear speech more accurately and improve the speech recognition performance. have.

Speech Recognition, Noise Reduction, Noise Adaptation, Model-Based Approach

Description

Noise reduction device and method {METHOD AND APPARATUS FOR REDUCTION OF NOISE}

The present invention relates to an apparatus and method for removing noise, and more particularly, to a method and apparatus for estimating clean speech from which dynamic noise has been removed from an input voice including noise for speech recognition in a noise environment.

The present invention is derived from the research conducted as part of the IT growth engine technology development project of the Ministry of Knowledge Economy and the Ministry of Information and Telecommunication Research and Development. [Task management number: 2006-S-036-03, Task name: Large-capacity interactive distribution for the new growth engine industry] Development of processing voice interface technology].

As is well known, the biggest obstacle for the commercialization of speech recognition technology is the degradation of recognition performance due to noise. It is not uncommon for speech recognition systems with near perfect performance in clean noise environments to noticeably degrade recognition performance in real environments with noise.

To address this problem, various approaches have been proposed so far, typically, signal-processing based spectral enhancement and statistical-model-based model adaptation. There is this.

The spectral enhancement technique estimates the noise spectrum in the absence of speech and subtracts this value from the spectrum of the noise-added input signal.The spectral subtraction and decision-directed methods The Wiener filter with the decision directed technique is the most widely used. Decision-oriented based Warner filters have been adopted as the ETSI Advanced Front-End (AFE) standard in a two-stage extended form and are known to be very effective in eliminating stationary noise.

Statistical model-based model adaptation technique converts acoustic models (generally hidden markov models) to noise conditions instead of modifying the input noise signal. The typical model is PMC (Parallel Model Compensation). . The PMC expresses clean speech and noise in different models and combines the two models to model mixed speech. It shows better performance than other techniques in the noise environment included in the noise model.

As described above, according to the prior art, various noise reduction techniques have been proposed to improve speech recognition performance in the presence of noise, but performance degradation of speech recognition still occurs, and dynamic noise removal efficiency is static in a noise environment. There is a problem in that it does not guarantee satisfactory voice recognition performance because it is inferior to noise. For example, the warner filter has a disadvantage in that performance improvement is limited in an environment in which a noise change is severe and an environment in which various noises are mixed.

The present invention has been proposed to solve the problems of the prior art, and improves the noise removal efficiency in an environment in which the noise change is severe and in the mixed environment of various noises, thereby improving the speech recognition performance by more accurately estimating the clean voice. A noise canceling device and method are provided.

As a first aspect of the present invention, a noise canceling apparatus includes a noise estimator for estimating a noise component included in an input speech signal, a post probability estimator for estimating a posterior probability value from the estimated noise component, and an estimation; A noise parameter adaptation unit adapted to adapt the noise Gaussian mixture model to the input speech signal based on the posterior probability value, the noise Gaussian mixture model, the speech Gaussian mixture model, and the input speech signal as inputs A voice / noise separation unit capable of separating the signal primarily, and a noise removing unit capable of removing residual noise components remaining in the separated voice signal.

Here, the noise estimator estimates the noise component using an Improved Minima Controlled Recursive Average (IMCRA) method.

The noise parameter adaptor adjusts a weight value of each Gaussian mixture of the noise Gaussian mixture model in proportion to the posterior probability value obtained by the posterior probability estimator.

The voice / noise separator separates the noise and the voice signal using a soft masking technique.

The noise canceller removes the residual noise component by using a decision directed warner filter.

According to a second aspect of the present invention, a noise removing method includes estimating a noise component included in an input speech signal, estimating a post-probability value from the estimated noise component, and based on the estimated post-probability value. Adapting a mixed model to the input speech signal, first separating the noise Gaussian mixed model and the speech Gaussian mixed model and the input speech signal as inputs, and separating the noise and the speech signal primarily; Removing residual noise components remaining in the voice signal.

In the estimating of the noise component, the noise component is estimated using an IMCRA method.

The adapting to the input speech signal adjusts a weight value of each Gaussian mixture of the noise Gaussian mixture model in proportion to the posterior probability value obtained by estimating the posterior probability value.

In the first step of separating, the noise and the voice signal are separated using a soft masking technique.

The step of removing the residual noise component removes the residual noise component using a decision directed based Warner filter.

According to the present invention, in order to improve noise rejection in environments with high noise fluctuations and various noises, voice / noise separation techniques such as soft masking techniques are used to enhance the separation of voice and noise, and To compensate for the limitation of Gaussian mixture model in modeling the noise component of the input signal, the noise adaptation technique is used to improve the speech recognition performance by estimating the clean speech more accurately.

Hereinafter, some embodiments of the present invention will be described in detail with reference to the accompanying drawings. In addition, in describing the present invention, when it is determined that the detailed description of the related known configuration or function may obscure the gist of the present invention, the detailed description thereof will be omitted.

1 is a block diagram of an apparatus for removing noise in accordance with an embodiment of the present invention.

As described above, the noise canceller of the present invention includes a noise estimator 130 capable of estimating a noise component included in an input speech signal, and a post probability estimator capable of estimating a posterior probability value from the estimated noise component. 140), a noise parameter adaptation unit 150 capable of adapting the noise Gaussian mixture model 110 to the input speech signal based on the estimated posterior probability value, an adaptive noise Gaussian mixture model 110 and a speech Gaussian mixture model ( 120) and a voice / noise separation unit 160 for separating noise and a voice signal primarily by inputting an input voice signal, and a noise removing unit for removing residual noise components remaining in the separated voice signal. 170, and the like.

A noise removing process by the noise removing device according to the embodiment of the present invention configured as described above will be described with reference to FIGS. 1 and 2.

First, the noise estimator 130 obtains an estimate of an approximate noise component from an input speech signal in which noise and speech are mixed (S205). For noise estimation, for example, IMCRA or the like can be used. The estimated noise components are then used to adapt the noise Gaussian mixture model 110 to the current noise conditions.

The posterior probability estimator 140 obtains a posterior probability value of the noise component estimated by the noise estimator 130 in the noise Gaussian mixture model 110 (S207). This may be expressed as Equation 1 below.

은 잡음 추정부(130)에서 추정한 잡음성분이며, k는 해당되는 잡음 가우시안 혼합 모델(110)의 가우시안 혼합값이고,

는

가 k에서 갖는 사후 확률값이다.From above

Is a noise component estimated by the noise estimator 130, k is a Gaussian mixture value of the corresponding noise Gaussian mixture model 110,

Is

Is the posterior probability value at k.

The noise Gaussian mixture model 110 is pre-trained (S201) using a noise database collected in a real environment (home environment, automobile, exhibition hall, etc.) where a speech recognition system is used, and models the noise of the environment in which the system is used. do. The voice Gaussian mixture model 120 is trained from a speech database collected in a noise-free environment (S203) to model clean speech from which noise is removed. The noise Gaussian mixture model 110 and the voice Gaussian mixture model 120 can be described as in Equation 2 below assuming a Gaussian mixture model composed of K mixtures.

The noise parameter adaptor 150 adjusts a weight value of each Gaussian mixture of the noise Gaussian mixture model 110 in proportion to the posterior probability value obtained by the posterior probability estimator 140. That is, it is assumed that a Gaussian having a high post probability value is better modeled as a noise component in the current state, and the value of the Gaussian mixture is adjusted in proportion to the post probability value (S209).

The voice / noise separation unit 160 inputs an input voice signal and a noise Gaussian mixture model 110 adapted to the current noise environment by the voice Gaussian mixture model 120 and the noise parameter adaptation unit 150. It performs primary function of separating noise. For example, a voice / noise separation technique such as a soft masking technique may be used (S211).

Since there may be residual noise components in the components separated by voice from the voice / noise separation unit 160, the noise removing unit 170 removes the residual noise components remaining in the first separated voice signal. Estimate the clean voice of (S213). The noise canceller 170 estimates a noise spectrum in a section where voice is not present using, for example, a decision-oriented warner filter, and then subtracts the value from the spectrum of the input signal to which the noise is added. It is available.

This estimated clean voice is used as input to the real voice recognition system.

The noise reduction method according to the present invention can be created by a computer program. The code and code segments that make up this computer program can be easily deduced by a computer programmer in the field. In addition, the computer program is stored in a computer readable media, and read and executed by a computer to implement a noise canceling method. The information storage medium includes a magnetic recording medium, an optical recording medium and a carrier wave medium.

So far, the present invention has been described with reference to some embodiments thereof. Those skilled in the art will appreciate that the present invention can be implemented in a modified form without departing from the essential features of the present invention. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the scope will be construed as being included in the present invention.

1 is a block diagram of an apparatus for removing noise in accordance with an embodiment of the present invention.

2 is a flowchart illustrating a noise removing method according to an embodiment of the present invention.

<Explanation of symbols for the main parts of the drawings>

110: noise Gaussian mixed model 120: voice Gaussian mixed model

130: noise estimation unit 140: post probability estimation unit

150: noise parameter adaptation unit 160: voice / noise separation unit

170: noise canceling unit

Claims (10)

delete A noise estimator for estimating a noise component included in an input speech signal; A post probability estimator for estimating a post probability value from the estimated noise component; A noise parameter adaptor adapted to adapt a noise Gaussian mixture model to the input speech signal based on the estimated posterior probability value; A speech / noise separation unit capable of separating the noise and the speech signal primarily by using the adapted noise Gaussian mixture model and the speech Gaussian mixture model and the input speech signal; And a noise removing unit for removing residual noise components remaining in the separated speech signal. The noise estimating unit estimates the noise component using an Improved Minima Controlled Recursive Average (IMCRA) method. Noise Canceling Device. A noise estimator for estimating a noise component included in an input speech signal; A post probability estimator for estimating a post probability value from the estimated noise component; A noise parameter adaptor adapted to adapt a noise Gaussian mixture model to the input speech signal based on the estimated posterior probability value; A speech / noise separation unit capable of separating the noise and the speech signal primarily by using the adapted noise Gaussian mixture model and the speech Gaussian mixture model and the input speech signal; And a noise removing unit for removing residual noise components remaining in the separated speech signal. The noise parameter adaptor adjusts a weight value of each Gaussian mixture of the noise Gaussian mixture model in proportion to the posterior probability value obtained by the posterior probability estimator. Noise Canceling Device. A noise estimator for estimating a noise component included in an input speech signal; A post probability estimator for estimating a post probability value from the estimated noise component; A noise parameter adaptor adapted to adapt a noise Gaussian mixture model to the input speech signal based on the estimated posterior probability value; A speech / noise separation unit capable of separating the noise and the speech signal primarily by using the adapted noise Gaussian mixture model and the speech Gaussian mixture model and the input speech signal; And a noise removing unit for removing residual noise components remaining in the separated speech signal. The voice / noise separation unit separates the noise and the voice signal using a soft masking technique. Noise Canceling Device. A noise estimator for estimating a noise component included in an input speech signal; A post probability estimator for estimating a post probability value from the estimated noise component; A noise parameter adaptor adapted to adapt a noise Gaussian mixture model to the input speech signal based on the estimated posterior probability value; A speech / noise separation unit capable of separating the noise and the speech signal primarily by using the adapted noise Gaussian mixture model and the speech Gaussian mixture model and the input speech signal; And a noise removing unit for removing residual noise components remaining in the separated speech signal. The noise removing unit removes the residual noise component by using a decision direction based Warner filter. Noise Canceling Device. delete Estimating a noise component included in an input speech signal; Estimating a posterior probability value from the estimated noise component; Adapting a noise Gaussian mixture model to the input speech signal based on the estimated posterior probability value; Firstly separating the noise and the speech signal using the adapted noise Gaussian mixture model, the speech Gaussian mixture model, and the input speech signal; Removing residual noise components remaining in the separated speech signal; The estimating of the noise component may include estimating the noise component using an Improved Minima Controlled Recursive Average (IMCRA) method. Noise Reduction Method. Estimating a noise component included in an input speech signal; Estimating a posterior probability value from the estimated noise component; Adapting a noise Gaussian mixture model to the input speech signal based on the estimated posterior probability value; Firstly separating the noise and the speech signal using the adapted noise Gaussian mixture model, the speech Gaussian mixture model, and the input speech signal; Removing residual noise components remaining in the separated speech signal; The adapting the input speech signal may include adjusting a weight value of each Gaussian mixture of the noise Gaussian mixture model in proportion to the posterior probability value obtained by estimating the posterior probability value. Noise Reduction Method. Estimating a noise component included in an input speech signal; Estimating a posterior probability value from the estimated noise component; Adapting a noise Gaussian mixture model to the input speech signal based on the estimated posterior probability value; Firstly separating the noise and the speech signal using the adapted noise Gaussian mixture model, the speech Gaussian mixture model, and the input speech signal; Removing residual noise components remaining in the separated speech signal; The first step of separating may include separating the noise and the voice signal using a soft masking technique. Noise Reduction Method. Estimating a noise component included in an input speech signal; Estimating a posterior probability value from the estimated noise component; Adapting a noise Gaussian mixture model to the input speech signal based on the estimated posterior probability value; Firstly separating the noise and the speech signal using the adapted noise Gaussian mixture model, the speech Gaussian mixture model, and the input speech signal; Removing residual noise components remaining in the separated speech signal; The removing of the residual noise component may include removing the residual noise component using a decision directed based Warner filter. Noise Reduction Method.

Images