Model-based Single-microphone Speech Separation Using Conditional Random Fields

Single-microphone speech separation requires to reconstruct two or more sources from only one speech mixture. It can serve as the front-end for speech applications that demand for robustness against interfering signals, such as information extraction from sound streams of multimedia. As an extreme case of under-determined source separation problem, a unique solution for source reconstruction is unlikely to be achieved, but the most probable source observations can be obtained through statistical inference given their prior information in a statistical model-based setting.;The performance of statistical model-based methods has been progressively improved by the use of graphical models to organize the prior information. In this thesis, the performance of the exact and the approximated statistical inference algorithms on single-microphone speech separation with factorial Hidden Markov models (HMM) are evaluated in terms of speech quality and computational complexity. The important role of state transitions in the source models is also investigated.;Model mis-specification is a major problem in model-based speech separation. These mis-specifications are caused by various factors, including limited amount of training data and finite number of acoustic states. Compared with generative approach such as factorial HMM, direct models like conditional random fields (CRF) are considered to be more robust to model mis-specification due to the inherent discrimination ability. In this thesis, the application of conditional random field (CRF) for single-microphone speech separation is investigated. The posterior probabilities of acoustic states given the mixture, which are essential to minimum mean-square error estimation of the sources, are modeled in a maximum entropy probability distribution. The performance of CRF formulations is further improved with a large-margin approach of parameter estimation.;Experimental results confirm that CRF formulations achieve the improved objective quality measures and automatic speech recognition accuracy of the reconstructed sources, especially when the sources are competing with similar signal-to-signal ratio. Even with a simplified CRF formulation, the performance is still comparable to factorial HMM.