Audio Source Separation Based on Residual Reprojection

조충상,김재우,이상근 한국전자통신연구원 2015 ETRI Journal Vol.37 No.4

This paper describes an audio source separation that is based on nonnegative matrix factorization (NMF) and expectation maximization (EM). For stable and high-performance separation, an effective auxiliary source separation that extracts source residuals and reprojects them onto proper sources is proposed by taking into account an ambiguous region among sources and a source’s refinement. Specifically, an additional NMF (model) is designed for the ambiguous region — whose elements are not easily represented by any existing or predefined NMFs of the sources. The residual signal can be extracted by inserting the aforementioned model into the NMF-EM-based audio separation. Then, it is refined by the weighted parameters of the separation and reprojected onto the separated sources. Experimental results demonstrate that the proposed scheme (outlined above) is more stable and outperforms existing algorithms by, on average, 4.4 dB in terms of the source distortion ratio.

인접 배치된 스테레오 무지향성 마이크로폰 환경에서 양이간 강도차를 활용한 음원 분리 기법

전찬준(Chan Jun Chun),정석희(Seok Hee Jeong),김홍국(Hong Kook Kim) 대한전자공학회 2013 전자공학회논문지 Vol.50 No.12

Flexible Nonlinear Learning for Source Separation

Park, Seung-Jin The Korean Institute of Electrical Engineers 2000 Journal of KIEE Vol.10 No.1

Source separation is a statistical method, the goal of which is to separate the linear instantaneous mixtures of statistically independent sources without resorting to any prior knowledge. This paper addresses a source separation algorithm which is able to separate the mixtures of sub- and super-Gaussian sources. The nonlinear function in the proposed algorithm is derived from the generalized Gaussian distribution that is a set of distributions parameterized by a real positive number (Gaussian exponent). Based on the relationship between the kurtosis and the Gaussian exponent, we present a simple and efficient way of selecting proper nonlinear functions for source separation. Useful behavior of the proposed method is demonstrated by computer simulations.

Multi-channel Speech Enhancement Using Blind Source Separation and Cross-channel Wiener Filtering

Jang, Gil-Jin,Choi, Chang-Kyu,Lee, Yong-Beom,Kim, Jeong-Su,Kim, Sang-Ryong The Acoustical Society of Korea 2004 韓國音響學會誌 Vol.23 No.e2

Despite abundant research outcomes of blind source separation (BSS) in many types of simulated environments, their performances are still not satisfactory to be applied to the real environments. The major obstacle may seem the finite filter length of the assumed mixing model and the nonlinear sensor noises. This paper presents a two-step speech enhancement method with multiple microphone inputs. The first step performs a frequency-domain BSS algorithm to produce multiple outputs without any prior knowledge of the mixed source signals. The second step further removes the remaining cross-channel interference by a spectral cancellation approach using a probabilistic source absence/presence detection technique. The desired primary source is detected every frame of the signal, and the secondary source is estimated in the power spectral domain using the other BSS output as a reference interfering source. Then the estimated secondary source is subtracted to reduce the cross-channel interference. Our experimental results show good separation enhancement performances on the real recordings of speech and music signals compared to the conventional BSS methods.

멀티채널 비음수 행렬분해와 정규화된 공간 공분산 행렬을 이용한 미결정 블라인드 소스 분리

오순묵,김정한 한국음향학회 2020 韓國音響學會誌 Vol.39 No.2

This paper solves the problem in underdetermined convolutive mixture by improving the disadvantages of the multichannel nonnegative matrix factorization technique widely used in blind source separation. In conventional researches based on Spatial Covariance Matrix (SCM), each element composed of values such as power gain of single channel and correlation tends to degrade the quality of the separated sources due to high variance. In this paper, level and frequency normalization is performed to effectively cluster the estimated sources. Therefore, we propose a novel SCM and an effective distance function for cluster pairs. In this paper, the proposed SCM is used for the initialization of the spatial model and used for hierarchical agglomerative clustering in the bottom-up approach. The proposed algorithm was experimented using the ‘Signal Separation Evaluation Campaign 2008 development dataset’. As a result, the improvement in most of the performance indicators was confirmed by utilizing the ‘Blind Source Separation Eval toolbox’, an objective source separation quality verification tool, and especially the performance superiority of the typical SDR of 1 dB to 3.5 dB was verified. 본 논문은 블라인드 소스 분리 분야에서 널리 사용되는 멀티채널 비음수 행렬 분해 기법의 단점을 개선하여 미결정 복잡한 혼합 환경에서 문제를 해결한다. 공간 공분산 행렬에 기반을 둔 기존의 연구들에서, 단일 채널의 파워게 인 및 상관관계와 같은 값으로 구성된 행렬의 각 요소는 높은 분산으로 인해 분리된 소스의 품질을 저하시키는 경향이 있다. 이 논문에서는 추정된 소스들을 효과적으로 클러스터링하기 위해 레벨 및 주파수 정규화를 수행한다. 따라서 새 로운 공간 공분산 행렬 및 효과적인 클러스터 쌍별 거리함수를 제안한다. 본 논문에서는 제안된 행렬을 공간 모델의 초기화에 활용하여 공간 모델의 향상된 추정과 이를 바탕으로 상향식 접근법에서의 계층적 응집 클러스터링에 활용함 으로써 분리된 음원의 품질을 향상시켰다. 제안된 알고리즘은 ‘Signal Separation Evaluation Campaign 2008 development dataset’을 활용하여 실험을 하였다. 그 결과 객관적인 소스 분리 품질 검증 도구인 ‘Blind Source Separation Eval toolbox’를 활용하여 대부분의 성능향상지표에서의 향상을 확인하였으며, 특히 대표적인 수치인 SDR의 1 dB ~ 3.5 dB 정도의 성능우위를 검증하였다.

유사-가능도 최대화를 통한 가우시안 프로세스 기반 음원분리

박선호(Sunho Park),최승진(Seungjin Choi) 한국정보과학회 2008 정보과학회논문지 : 소프트웨어 및 응용 Vol.35 No.7

본 논문에서는 각 음원이 시간적 구조를 가졌을 경우 음원들을 분리해내는 확률적 음원분리 방법을 제안한다. 이를 위해 각 음원의 시간적 구조를 가우시안 프로세스(Gaussian process)로 모델링하고 기존의 음원분리 문제를 유사-가능도 최대화 문제(pseudo-likelihood maximization)로 공식화한다. 본 알고리즘을 통해 얻어진 데이타의 유사-가능도는 정규 분포이며 이는 가우시안 프로세스 회귀방법(Gaussian process regression)을 통해 쉽게 계산이 가능하다. 음원분리의 역혼합 행렬은 경도(gradient) 기반 최적화 기법을 통해 데이타의 유사-가능도를 최대화 하는 해를 찾음으로써 구해진다. 여러 실험을 통하여 제안 알고리듬이 몇 가지 특정 상황에서 기존의 분리 알고리듬들에 비해 우수한 성능을 보임을 확인 할 수 있다. In this paper we present a probabilistic method for source separation in the case where each source has a certain temporal structure. We tackle the problem of source separation by maximum pseudo-likelihood estimation, representing the latent function which characterizes the temporal structure of each source by a random process with a Gaussian prior. The resulting pseudo-likelihood of the data is Gaussian, determined by a mixing matrix as well as by the predictive mean and covariance matrix that can easily be computed by Gaussian process (GP) regression. Gradient-based optimization is applied to estimate the demixing matrix through maximizing the log-pseudo-likelihood of the data. umerical experiments confirm the useful behavior of our method, compared to existing source separation methods.

반사음이 존재하는 양귀 모델의 음원분리에 관한 연구

이채봉,Lee, Chai-Bong 한국융합신호처리학회 2014 융합신호처리학회 논문지 (JISPS) Vol.15 No.3

두 개의 입력소자에 의한 음원방향 및 분리방법으로서는 연산량이 적고, 음원분리 성능이 높은 주파수 양귀 모델(Frequency Domain Binaural Model : FDBM)이 있다. FDBM은 주파수 영역에서 양귀간 위상차(Interaural Phase Difference : IPD) 및 양귀간 레벨차(Interaural Level Difference : ILD)를 구하여 음향신호가 오는 방향과 음원의 분리처리를 한다. 그러나 실제 환경에서는 반사음의 문제가 되고 있다. 이러한 반사음에 의한 영향을 줄이기 위하여 선행음 효과에 의한 직접음의 음상정위를 모의하여 초기 도착음을 검출하고 직접음이 오는 방향과 음원분리 방법을 제시하였다. 제시한 방법을 이용하여 음원방향 추정 및 분리에 대한 성능을 시뮬레이션으로 검토하였다. 그 결과, 방향추정은 음원이 오는 방향에서 ${\pm}10%$의 범위로 집중되어 음원의 방향과 가까운 값으로 추정되었다, 반사음이 존재하는 경우의 음원분리는 기존의 FDBM에 비하여 코히런스(Coherence), 음성품질 지각평가 PESQ(Perceptual Evaluation of Speech Quality : PESQ)가 높고, 정면에서의 지향특성 감쇠량이 작아 분리의 정도가 개선됨을 나타내었다. 그러나 반사음이 존재하지 않는 경우는 분리 정도가 낮았다. For Sound source direction and separation method, Frequency Domain Binaural Model(FDBM) shows low computational cost and high performance for sound source separation. This method performs sound source orientation and separation by obtaining the Interaural Phase Difference(IPD) and Interaural Level Difference(ILD) in frequency domain. But the problem of reflection occurs in practical environment. To reduce this reflection, a method to simulate the sound localization of a direct sound, to detect the initial arriving sound, to check the direction of the sound, and to separate the sound is presented. Simulation results show that the direction is estimated to lie close within 10% from the sound source and, in the presence of the reflection, the level of the separation of the sound source is improved by higher Coherence and PESQ(Perceptual Evaluation of Speech Quality) and by lower directional damping than those of the existing FDBM. In case of no reflection, the degree of separation was low.

A simple iterative independent component analysis algorithm for vibration source signal identification of complex structures

Lee, Dong-Sup,Cho, Dae-Seung,Kim, Kookhyun,Jeon, Jae-Jin,Jung, Woo-Jin,Kang, Myeng-Hwan,Kim, Jae-Ho The Society of Naval Architects of Korea 2015 International Journal of Naval Architecture and Oc Vol.7 No.1

Independent Component Analysis (ICA), one of the blind source separation methods, can be applied for extracting unknown source signals only from received signals. This is accomplished by finding statistical independence of signal mixtures and has been successfully applied to myriad fields such as medical science, image processing, and numerous others. Nevertheless, there are inherent problems that have been reported when using this technique: instability and invalid ordering of separated signals, particularly when using a conventional ICA technique in vibratory source signal identification of complex structures. In this study, a simple iterative algorithm of the conventional ICA has been proposed to mitigate these problems. The proposed method to extract more stable source signals having valid order includes an iterative and reordering process of extracted mixing matrix to reconstruct finally converged source signals, referring to the magnitudes of correlation coefficients between the intermediately separated signals and the signals measured on or nearby sources. In order to review the problems of the conventional ICA technique and to validate the proposed method, numerical analyses have been carried out for a virtual response model and a 30 m class submarine model. Moreover, in order to investigate applicability of the proposed method to real problem of complex structure, an experiment has been carried out for a scaled submarine mockup. The results show that the proposed method could resolve the inherent problems of a conventional ICA technique.

A simple iterative independent component analysis algorithm for vibration source signal identification of complex structures

이동섭,조대승,김국현,전재진,정우진,강명환,김재호 대한조선학회 2015 International Journal of Naval Architecture and Oc Vol.7 No.1

Independent Component Analysis (ICA), one of the blind source separation methods, can be applied for extracting unknown source signals only from received signals. This is accomplished by finding statistical independence of signal mixtures and has been successfully applied to myriad fields such as medical science, image processing, and numerous others. Nevertheless, there are inherent problems that have been reported when using this technique: instability and invalid ordering of separated signals, particularly when using a conventional ICA technique in vibratory source signal identification of complex structures. In this study, a simple iterative algorithm of the conventional ICA has been proposed to mitigate these problems. The proposed method to extract more stable source signals having valid order includes an iterative and reordering process of extracted mixing matrix to reconstruct finally converged source signals, referring to the magnitudes of correlation coefficients between the intermediately separated signals and the signals measured on or nearby sources. In order to review the problems of the conventional ICA technique and to validate the proposed method, numerical analyses have been carried out for a virtual response model and a 30 m class submarine model. Moreover, in order to investigate applicability of the proposed method to real problem of complex structure, an experiment has been carried out for a scaled submarine mockup. The results show that the proposed method could resolve the inherent problems of a conventional ICA technique.

주파수 특성 기저벡터 학습을 통한 특정화자 음성 복원 (pp.179-186)

박선호(Sunho Park),유지호(Jiho Yoo),최승진(Seungjin Choi) 한국정보과학회 2009 정보과학회논문지 : 소프트웨어 및 응용 Vol.36 No.3

본 논문에서는 학습이 가능한 특정화자의 발화음성이 있는 경우, 잡음과 반향이 있는 실 환경에서의 스테레오 마이크로폰을 이용한 특정화자 음성복원 알고리즘을 제안한다. 이를 위해 반향이 있는 환경에서 음원들을 분리하는 다중경로 암묵음원분리(convolutive blind source separation, CBSS)와 이의 후처리 방법을 결합함으로써, 잡음이 섞인 다중경로 신호로부터 잡음과 반향을 제거하고 특정화자의 음성만을 복원하는 시스템을 제시한다. 즉, 비음수 행렬분해(non-negative matrix factorization, NMF) 방법을 이용하여 특정화자의 학습음성으로부터 주파수 특성을 보존하는 기저벡터들을 학습하고, 이 기저벡터들에 기반 한 두 단계의 후처리 기법들을 제안한다. 먼저 본 시스템의 중간단계인 CBSS가 다중경로 신호를 입력받아 독립음원들을(두 채널) 출력하고, 이 두 채널 중 특정화자의 음성에 보다 가까운 채널을 자동적으로 선택한다(채널선택 단계). 이후 앞서 선택된 채널의 신호에 남아있는 잡음과 다른 방해음원(interference source)을 제거하여 특정화자의 음성만을 복원, 최종적으로 잡음과 반향이 제거된 특정화자의 음성을 복원한다(복원 단계). 이 두 후처리 단계 모두 특정화자 음성으로부터 학습한 기저벡터들을 이용하여 동작하므로 특정화자의 음성이 가지는 고유의 주파수 특성 정보를 효율적으로 음성복원에 이용 할 수 있다. 이로써 본 논문은 CBSS에 음원의 사전정보를 결합하는 방법을 제시하고 기존의 CBSS의 분리 결과를 향상시키는 동시에 특정화자만의 음성을 복원하는 시스템을 제안한다. 실험을 통하여 본 제안 방법이 잡음과 반향 환경에서 특정화자의 음성을 성공적으로 복원함을 확인할 수 있다. This paper proposes a target speech extraction which restores speech signal of a target speaker form noisy convolutive mixture of speech and an interference source. We assume that the target speaker is known and his/her utterances are available in the training time. Incorporating the additional information extracted from the training utterances into the separation, we combine convolutive blind source separation(CBSS) and non-negative decomposition techniques, e.g., probabilistic latent variable model. The nonnegative decomposition is used to learn a set of bases from the spectrogram of the training utterances, where the bases represent the spectral information corresponding to the target speaker. Based on the learned spectral bases, our method provides two postprocessing steps for CBSS. Channel selection step finds a desirable output channel from CBSS, which dominantly contains the target speech. Reconstruct step recovers the original spectrogram of the target speech from the selected output channel so that the remained interference source and background noise are suppressed. Experimental results show that our method substantially improves the separation results of CBSS and, as a result, successfully recovers the target speech.