KR101897242B1 - A method for enhancing quality of speech including noise - Google Patents

Abstract

An apparatus for improving voice signal quality is disclosed. The apparatus for enhancing speech signal quality according to an embodiment of the present invention includes an a priori SNR estimator for estimating a priori signal to noise ratio (SNR) for a speech signal including noise, a first Wiener filter gain obtained using the SNR And a non-noise matrix factorization processor for obtaining a non-noise matrix factorization gain by performing a non-noise matrix factorization on a voice signal including a noise applied with the first Wiener filter gain, The filter processing unit obtains the second Wiener filter gain used for acquiring the second voice signal using the non-noise matrix factorization gain and the a priori SNR.

Description

[0001] The present invention relates to a method for enhancing the quality of speech including noise,

The present invention relates to a method for improving the sound quality of a voice including noise. More particularly, to a method for improving the sound quality of a voice by using a non-numerical matrix factorization and a Wiener filter.

Speech enhancement techniques such as the Wiener filter and the minimum mean square error (MMSE) estimation, which are commonly used to improve the quality of noise speech, are used to estimate noise power spectral density (PSD) To-noise ratio (SNR) with respect to a priori signal due to the presence of the signal. Such statistical based speech enhancement techniques such as Wiener filter are known to exhibit relatively good performance in a stationary noise environment.

However, in a non-statoinary noise environment such as a babble, the accuracy of the noise PSD and the a priori SNR estimation becomes poor, which causes a residual noise problem after the speech enhancement processing.

Non-negative matrix factorization (NMF), which is being studied as an alternative method for improving speech, is to extract the correct noise from the observed speech in order to obtain the basis of noise by using the basis of non- Although the detection of the interval is required, this is also a challenge that has been undergoing much research for decades.

As a result, it is necessary to develop a non-speech matrix factorization-based speech enhancement method that improves speech quality only at the basis of speech without basis of noise.

Prior Art Document 1: M. B. Trawicki, M. T. Johnson, Distributed multichannel speech enhancement with minimum mean-square error, short-time spectral amplitude, log-spectral amplitude, and spectral phase estimation. Signal Processing. 92 (2) (2012) 345-356. Prior Art Document 2: W. Lee, J. H. Song, J. H. Chang, Minima-controlled speech presence uncertainty tracking method for speech enhancement. Signal Processing. 91 (1) (2011) 155-161. Prior Art Document 3: P. C. Loizou, Speech Enhancement: Theory and Practice. Second Edition. CRC 2013. Prior Art Document 4: Speech Processing, Transmission and Quality Aspects (STQ); Distributed speech recognition; Advanced frontend feature extraction algorithm; Compression algorithms, Tech. Rep. ETSI ES 202 050 V1.1.5 2007. Prior Art Document 5: S. J. Lee, B. O. Kang, H. Jung, Y. Lee, H. S. Kim, Statistical Model-Based Noise Reduction Approach to Speech Recognition in Speech Recognition. ETRI Journal. 32 (5) (2010) 801-809. Prior Art Document 6: S. Mirzaei, H. V. Hamme, Y. Norouzi, Blind audio source counting and separation of anechoic mixtures using the multichannel complex NMF framework. Signal Processing. 115 (2015) 27-37. Prior Art Document 7: Y. Xu, G. Bao, X. Xu, Z. Ye, Single-channel speech separation using sequential discriminative dictionary learning. Signal Processing. 106 (2015) 134-140. Prior Art Document 8: F. Weninger, J. Le Roux, J. R. Hershey, S. Watanabe, Discriminative NMF and its application to single-channel source separation. Proc. Interspeech. Singapore (2014) 865-869. Prior Art Document 9: N. Mohammadiha, P. Smaragdis, A. Leijon, Supervised and Unsupervised Speech Enhancement Using Nonnegative Matrix Factorization. IEEE Trans. Audio, Speech, and Language Process. 21 (10) (2013) 2140-2151. Prior Art Document 10: C. Joder, F. Weninger, F. Eyben, D. Virette, B. Schuller, Real-time speech separation by semi-supervised nonnegative matrix factorization. Proc. International Conference on Latent Variable Analysis and Signal Separation. Tel Aviv. Israel (2012) 322-329. Prior Art Document 11: H. Hu, A. Krasoulis, M. Lutman, S. Bleeck, Development of a real time sparse non-negative matrix factorization module for cochlear implants by using xPC target. Sensors. 13 (10) (2013) 13861? 13878. Prior Art Document 12: P. Hoyer, Non-negative sparse coding. Proc. IEEE Workshop on Neural Networks for Signal Processing. (2002) 557-565. Prior Art Document 13: ITU-T Recommendation P.862, Perceptual Evaluation of Speech Quality (PESQ), and Objective Method for End-End Speech Quality Assessment of Narrowband Telephone Networks and Speech Coders Feb. 2001. Prior Art Document 14: J. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, N. L. Dahlgren, V. Zue, TIMIT Acoustic Phonetic Continuous Speech Corpus (Linguistic Data Consortium, Philadelphia, 1993). Prior Art Document 15: A. Varga, H. J. M. Steenneken, M. Tomilson, D. Jones, The NOISEX-92 study on additive noise on automatic speech recognition (Documentation on NOISEX-92 CD-ROMs 1992).

An embodiment of the present invention is intended to improve the quality of a voice including noises by extending and improving the algorithm based on the factorization of a non-sound number matrix.

The apparatus for enhancing speech signal quality according to an embodiment of the present invention includes an a priori SNR estimator for estimating a a priori signal to noise ratio (SNR) for a speech signal including a noise, And a non-noise matrix factorization processing unit for obtaining a non-noise matrix factorization gain by performing a non-noise matrix factorization process on a voice signal including a noise applied with the first Wiener filter gain, And the Wiener filter gain generator obtains a second Wiener filter gain used for acquiring the second speech signal using the non-noise matrix factorization gain and a-priori SNR.

The speech quality enhancement method according to an embodiment of the present invention can improve the quality of the noise-including speech by expanding and improving the algorithm based on the factorization of the non-sound number matrix.

1 is a block diagram of a speech quality enhancement apparatus according to an embodiment of the present invention.
FIG. 2 is a graph illustrating Perceptual Evaluation of Speech Quality (PESQ) score obtained according to a method proposed in an embodiment of the present invention, according to a noise type.
Figure 3 shows spectral residual noise.
Figure 4 shows the spectrogram at each condition.
5 is a diagram showing PESQ scores for each method in different types of noise environments.
Figure 6 shows the results of a comparison of the method PR according to an embodiment of the present invention with Bayesian NMF (BNMF-HMM) (see prior art document 9) R4 combined with a Hidden Markov model (HMM).
FIG. 7 is a flowchart illustrating a method for improving speech quality using a factorizing factor filter Wiener filter according to an embodiment of the present invention. Referring to FIG.

Hereinafter, specific embodiments of the present invention will be described in detail with reference to the drawings. However, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. It will be understood by those skilled in the art that the present invention may be embodied in many other specific forms without departing from the spirit or essential characteristics thereof.

이라고 할 수 있다.Hereinafter, a method of improving the quality of a received voice signal is proposed. Specifically, a method for acquiring a voice easy to identify by separating / eliminating noise in a voice signal including noise is proposed. Herein, the speech signal obtained according to an embodiment of the present invention may be referred to as a target speech,

.

Hereinafter, a speech quality enhancement method and apparatus according to an embodiment of the present invention will be described with reference to FIG.

이 주파수 영역에서 부가적인 잡음

에 의해 ㅌ오염된다고 가정하면, 잡음이 포함된 음성

은 k번째 주파수 및 k번째 프레임에서 수학식 1로 표현된다.Initial voice signal

Additional noise in this frequency range

Assuming that the noise is contaminated by noise,

Is expressed by Equation (1) in the k- th frequency and the k- th frame.

은 목표 음성 추정을 시도하기 위해 수학식 2로 표현된다.Spectral gain for voice enhancement

Is expressed by Equation (2) to try target speech estimation.

은 DD 어프로치(decision-directed approach)에 기반한 잡음 분산 추정치에 의해 추정되는 선험적 SNR 추정치

에 의해 수학식 3과 같이 표현된다. DD 어프로치에 대한 상세한 설명은 선행기술문헌 3을 참고한다.In particular, the gain in the Wiener filter

0.0 > SNR estimate < / RTI > estimated by a noise variance estimate based on a decision-directed approach (DD)

Is expressed by Equation (3). For a detailed description of the DD approach, see Prior Art Document 3.

를 향상시키기 위한 것이다. 이를 설명하기 위해 이하에서 본 발명의 일 실시 예에 따른 음성 품질 향상 장치를 구체적으로 설명한다.The method for enhancing speech quality according to an embodiment of the present invention includes removing residual noise components remaining after applying the Wiener filter gain

. In order to explain the above, a speech quality enhancement apparatus according to an embodiment of the present invention will be described in detail.

1 is a block diagram of a speech quality enhancement apparatus according to an embodiment of the present invention.

The apparatus for enhancing voice quality 1 according to the embodiment of the present invention includes a priori SNR estimator 10, a winner filter gain generator 20, a non-noise matrix factorization processor 30 and a winner filter gain application unit 40 ).

를 추정한다. 구체적인 추정 방법은 선행기술문헌 4 및 5를 참고한다.The a priori SNR estimator 10 estimates a priori signal-to-noise ratio (SNR) based on the DD approach,

. Refer to prior art documents 4 and 5 for specific estimation methods.

The non-sound-number matrix factorization processor 30 estimates a non-sound performance factor-based gain in an online framework.

을 향상시키기 위하여, 비음수행렬 인수분해 처리부(30)는 전 대역 주파수를 통해

을 벡터

로 구성한다. 이는 수학식 4로 표현된다Using the non-numeric matrix factorization method

The non-sound number matrix factorization processing unit 30 performs a non-sound number matrix factorization processing

Vector

. This is expressed by Equation 4

는 사전에 학습된 기저행렬(pre-trained basis matrix)

및 활성화(activation) 벡터

을 이용하여 수학식 5와 같이 표현된다.Where T is the transpose operator. And

A pre-trained basis matrix < RTI ID = 0.0 >

And an activation vector

Is expressed as Equation (5).

는 모든 주파수에서 위너필터 게인

을 적용한 뒤 남아있는 잔여(residual) 잡음 성분(component)으로 구성된 벡터를 나타낸다. 이때,

는 임의의 목표 음성을 재구성하기 위한 범용 음성 데이터베이스로부터 학습(training)된다.From here

Is the gain of the Wiener filter at all frequencies

And a residual noise component remaining after applying the residual noise component. At this time,

Is trained from a general purpose speech database for reconstructing any target speech.

를 최소화하는

를 찾기 위한 것이다.

는 랜덤 초기화 된 후 오차 e 값이 최소화된 값으로 수렴할 때까지 반복적으로 업데이트 되어 추정된다. 이는 수학식 6으로 표현된다.Therefore, the non-tone number matrix factorization process is error-

To minimize

.

Is randomly initialized and then iteratively updated until the error e value converges to the minimized value. This is expressed by Equation (6).

및 나누기는 요소별 연산을 나타낸다. 그리고 마지막으로 비음수행렬 인수분해 처리부(30)는 위너 필터의 잔류 잡음을 감쇄시키기 위한 비음수행렬 인수분해 기반 게인

을 획득한다. 이는 수학식 7과 같이 표현된다.From here,

And division represent element-specific operations. Finally, the non-sound number matrix factorization processing unit 30 calculates a non-sound number matrix factorization based gain for attenuating the residual noise of the Wiener filter

. This is expressed by Equation (7).

은 분자가 0이 되는 것을 피하기 위한 최소값이다. 그리고, 획득된 비음수행렬 인수분해 기반 게인

는

을 향상시키기 위해 사용된다. 이는 수학식 8과 같이 표현된다. From here,

Is the minimum value to avoid zero molecules. Then, the obtained non-noise number matrix factorization-based gain

The

. ≪ / RTI > This is expressed by Equation (8).

및 선험적 SNR 추정치

를 모두 고려하여 업데이트된 새로운 위너 필터 게인

을 획득한다. In other words, the winner filter gain generator 20 generates the winner filter gain gain

And a priori SNR estimate

New Winner Filter Gain Updated

.

)에 따라 획득된 PESQ 스코어를 잡음 종류에 따라 도시한 그래프이다.FIG. 2 is a graph showing a sparse parameter of a method proposed in an embodiment of the present invention;

) According to the type of noise. ≪ tb >< TABLE >

(a) shows a case where the noise type is a babble, and (b) shows a case where the noise type is Gaussian noise.

As shown in FIG. 2, when comparing the PESQ score (a score obtained by scoring the sound quality from 1 point to 4.5 points to indicate that the sound quality is better as the score is higher), the prior art reference 13) It can be seen that the sparse parameter has a great influence on the sound quality improvement.

는 수학식 9와 같이 표현된다.The improved speech estimate < RTI ID = 0.0 >

Is expressed by Equation (9).

는 신호 변환부(미도시)에서 역 이산 푸리에 변환(DFT) 및 중첩합(overlap-and-add) 방법에 의해 시간-이산(time-discrete) 신호로 변환된다. 그리고, 추정된

는 선험적 SNR 추정부(10)에서

을 추정하는데 다시 이용된다.Then,

Discrete signal by an inverse discrete Fourier transform (DFT) and an overlap-and-add method in a signal converter (not shown). Then,

The a priori SNR estimator 10

Is estimated.

Figure 3 shows spectral residual noise.

In FIG. 3, the black region indicates the unfiltered noise included in the speech when the SNR is 5 in the frequency domain. Here, (a) is the babble noise and (b) is the Gaussian noise. The dark gray area represents residual noise after applying the Wiener filter gain to the noise containing the noise. And the light gray region represents the residual noise after applying the non-noise matrix factorization.

As shown in Fig. 3, the Wiener filter is effective in reducing noise. On the other hand, the remaining residual noise is further removed by non-tone matrix factorization using a previously learned speech basis.

In order to remove the noise, Euclidean distance-based Nominal Matrix factorization (UCL-NMF) and Kullback-Leibler divergence based non-numerical factorization (KL-NMF) can be used.

It may be more preferable to use the Euclidean distance based non-noise matrix factorization when using the non-sound number matrix factorization technique according to a specific embodiment of the present invention.

Because the voice quality enhancement method according to an embodiment of the present invention uses the voice basis only as a database and does not use a noise basis, the Euclidean distance-based non-voice matrix factorization, which tends to extract Gaussian noise, . ≪ / RTI > This is also shown in Table 1.

SNR (dB) Wiener Wiener + KL-NMF Wiener + UCL-NMF [20,15] 3.171 3.192 3.195 [10,5] 2.438 2.469 2.518 [0, -5] 1.644 1.679 1.805

Table 1 shows the PESQ scores of the bubble noise due to combination of different non-noise matrix factorization types and Wiener filters in the speech quality enhancement method according to an embodiment of the present invention.

Figure 4 shows the spectrogram at each condition.

에 따라 처리된 음성 신호를 나타내며, (d)는 선행기술문헌 11에 따른 non-negative sparse coding이 적용된 음성 신호를 나타내며, (e)는 본 발명의 일 실시 예에 따른 업데이트된 위너 필터 게인

에 따라 처리된 음성 신호를 나타낸다.Specifically, (a) represents a clean speech, (b) represents a voice including a background noise at 0 dB SNR, and (c)

(D) shows a speech signal to which non-negative sparse coding according to the prior art document 11 is applied, and (e) shows an updated Wiener filter gain according to an embodiment of the present invention.

Lt; / RTI >

As shown in FIG. 4, it can be seen that (e) according to an embodiment of the present invention is effective for improving speech quality as compared with the general speech quality improvement techniques (c) and (d).

5 is a diagram showing PESQ scores for each method in different types of noise environments.

In the case of non-negative sparse coding, R2 is a two-stage Mel-warped Wiener filter (see prior art document 4), R3 is a model In the case where the -based Wiener filter is applied (see Prior Art Document 5), PR represents a case where the non-noise matrix factorization Winner filter according to an embodiment of the present invention is applied.

As shown in FIG. 5, it can be seen that the method according to an embodiment of the present invention exhibits a better effect in all noise conditions except for the vehicle traveling noise condition.

Figure 6 shows the results of a comparison of the method PR according to an embodiment of the present invention with Bayesian NMF (BNMF-HMM) (see prior art document 9) R4 combined with a Hidden Markov model (HMM).

As shown in FIG. 6, it can be seen that the method according to an embodiment of the present invention has a better effect than R4 under all noise conditions.

FIG. 7 is a flowchart illustrating a method for improving speech quality using a factorizing factor filter Wiener filter according to an embodiment of the present invention. Referring to FIG.

The speech quality enhancement apparatus 1 according to an embodiment of the present invention acquires a speech signal including noise (S101). Here, the noise may be, for example, babble or Gaussian noise. The apparatus 1 for enhancing voice quality according to an embodiment of the present invention can extract only a voice by separating noise and voice. The extracted speech can be used for automatic speech recognition and a hearing aid.

또는 잡음이 포함된 음성 신호

로부터 선험적 SNR 추정치를 획득할 수 있다.The a-priori SNR estimator 10 acquires a priori SNR estimate for a speech signal including noise (S103). In a specific embodiment, the a-priori SNR estimator estimates a speech signal

Or a voice signal containing noise

Lt; RTI ID = 0.0 > SNR < / RTI >

The Wiener filter gain generator 20 acquires the first Wiener filter gain based on the obtained a-priori SNR estimate (S105). The first Wiener filter gain can be expressed as Equation (3).

The Wiener filter gain application unit 40 obtains a first enhanced speech signal in which the first Wiener filter gain is applied to the speech signal including the acquired noise (S107). Here, the first enhanced speech signal can be expressed as Equation (2).

The non-sound number matrix factorization processing unit 30 obtains the non-sound number matrix factorization gain by processing the non-sound number matrix factorization on the first enhanced speech signal to which the first Wiener filter gain is applied (S109). Specifically, the non-voice matrix factorization processor 30 acquires the non-voice matrix factorization gain using only the voice basis except for the noise base in the database learning.

The Wiener filter processing unit 20 obtains the updated second Wiener filter gain using the non-noise matrix factorization gain and the a-priori SNR estimate (S111). The updated second Wiener filter gain can be expressed as Equation (8).

The Wiener filter gain application unit 40 applies the second Wiener filter gain to the voice including the noise to output the second enhanced voice signal (S113). Here, the speech signal to which the second Wiener filter gain is applied can be expressed by Equation (9). The final output speech signal is converted into a time-discrete signal in the signal conversion unit.

The method for improving the voice signal quality according to an embodiment of the present invention can be implemented by a program. Specifically, it can be coded according to the above-described algorithm and implemented as a program. A program implementing the algorithm according to an embodiment of the present invention can be recorded on a storage medium.

As described above, the present invention has been described with reference to particular embodiments, such as specific elements, and limited embodiments and drawings. However, it should be understood that the present invention is not limited to the above- And various modifications and changes may be made thereto without departing from the scope of the present invention. Accordingly, the spirit of the present invention should not be construed as being limited to the embodiments described, and all of the equivalents or equivalents of the claims, as well as the following claims, fall within the scope of the spirit of the present invention .

Claims (6)

Obtaining a speech signal including noise;
Obtaining a priori signal to noise ratio (SNR) for a speech signal including noise;
Obtaining a first Wiener filter gain using the a-priori SNR;
Obtaining a first speech signal to which the first Wiener filter gain is applied to the speech signal including the noise;
Obtaining a non-sound number matrix factorization gain by factorizing a non-voice number matrix of the first voice signal;
Acquiring a second Wiener filter gain obtained by updating the first Wiener filter gain using the non-noise matrix factorization gain and the a priori signal-to-noise ratio; And
And applying the second Wiener filter gain to a speech signal including noise to obtain a second speech signal
How to improve voice signal quality. The method according to claim 1,
The step of factorizing the non-voice number matrix of the first speech signal
And machine learning the first speech signal using only the speech database
How to improve voice signal quality. 3. The method of claim 2,
The machine learning technique is based on the Euclidean distance-based non-note number matrix factorization (UCL-NMF)
How to improve voice signal quality. A priori SNR estimator for estimating a priori signal to noise ratio (SNR) for a speech signal including noise;
A Wiener filter gain generator for obtaining a first Wiener filter gain using the a priori SNR; And
And a non-sound number matrix factorization processor for obtaining a non-sound number matrix factorization gain by factorizing a non-sound number matrix of a first sound signal including noise to which the first Wiener filter gain is applied,
The Wiener filter gain generator may obtain a second Wiener filter gain that updates the first Wiener filter gain used for acquiring the second voice signal using the non-noise matrix factorization gain and the a priori SNR
A device for enhancing voice signal quality. 5. The method of claim 4,
Wherein the non-sound-number matrix factorization processor uses the speech database only to perform the machine learning of the first speech signal
A device for enhancing voice signal quality. 5. The method of claim 4,
The non-sound number matrix factorization processor may perform the machine learning of the first speech signal by factorizing the Euclidean distance-based non-
A device for enhancing voice signal quality.

Images