Font Size: a A A

Blind Separation Based On Source Model And Its Application

Posted on:2008-07-10Degree:DoctorType:Dissertation
Country:ChinaCandidate:Q H HuangFull Text:PDF
GTID:1118360215976850Subject:Pattern Recognition and Intelligent Systems
Abstract/Summary:PDF Full Text Request
Blind source separation (BSS) is one emerging area in modern signal processing. It has solid theoretical foundations and many potential applications. The assumptions on the instantaneous linear model of BSS are ideal. They have many limitations in practical applications. Therefore, the current research in domestic and foreign focuses on noisy mixing, nonstationary mixing and underdetermined mixing (the number of sensors is less than that of sources) etc. We do some research for some of the above problems.Noisy BSS is more appropriate to practical applications. So our work focuses on noisy BSS, that is, sources are separated from noisy observed signals. We not only investigate into the case where the number of observations is not less than that of sources, but also study single-channel BSS and recover two sources from only one observation. The temporal characteristics of source signals are exploited and some appropriate models are adopted to describe the temporal structure. All our separation algorithms are based on variational Bayesian learning. They integrate additive noise and temporally correlated information of sources to reduce noise and separate source signals. Finally, independent component analysis (ICA) is used to enhance noisy speech and more effective methods for noise shrinkage have been proposed.The main contributions of the dissertation are as follows:(1) We separate sources from noisy observed signals and consider noise reduction and source separation in an integrated manner. As a result, our proposed algorithms are robust in noisy environment.(2) The temporal structure of each source is exploited. Some appropriate models, i.e., autoregressive model, generalized autoregressive model and time-varying autoregressive model are adopted to describe the temporal information of source signal in order to improve the accuracy of separated results.(3) Single-channel BSS is a more challengeable problem that we investigate. Noisy single-channel mixture and autoregressive source model are combined to build a state-space model. In this way, source separation is transformed into state estimation. Variational Kalman smoothing can be adopted to recover two sources.(4) All our BSS algorithms are based on variational Bayesian learning due to its obvious advantage. They not only consider additive system noise but also exploit the source temporal information. Therefore they can avoid overfitting. Moreover, they quantificationally compare different models and select the model which is most appropriate to describe the available data.(5) ICA has an important application in speech enhancement. After ICA transform, the basis coefficients of speech signal are sparser. We choose appropriate probability density function (pdf) to model the sparsity and develop two more effective shrinkage algorithms.
Keywords/Search Tags:blind source separation, independent component analysis, variational Bayesian learning, autoregressive model, state space model, mixture of Gaussian model, variational Kalman smoothing, speech enhancement
PDF Full Text Request
Related items