Short-time Independent Component Analysis for blind separation of speech sources

Independent Component Analysis (ICA) has long been regarded as a powerful technique for speech source separation. In practice, however, speaker moving or reverberant environments may necessitate ICA to be implemented in short time intervals, which makes the fundamental assumption of sources' independence collapse in ICA. This leads to two important but often overlooked problems, namely: (1) excursion of global optimum from the desired solution and (2) diffusion of local optima in search of the de-mixing matrix. These two problems occur in most practical situations and greatly degrade the performance of the existing ICA algorithms.; Based on the insight into the effect on the aforementioned problems by the input sources as well as the mixing channel, three basic short time Local Optima Distribution (LOD) types are investigated. Information is derived from the characteristics of these LOD types for: (1) choosing simultaneous or sequential ICA algorithm; (2) shrinking feasible search region; and (3) producing possible initial points in search of the de-mixing matrix. As a result, the technique of LOD-based ICA is developed in this thesis to assign different procedures according to the LOD type of the observed short time mixtures. The analytical and simulation results demonstrated that more accurate de-mixing matrix estimation could be obtained; thereby producing improved separation performance.; Among all the three LOD types, the Dominant LOD manifests to be with comparatively higher efficiency in yielding accurate separation performance. The production mechanism of the Dominant LOD indicates that higher energy ratio of sources helps to build this type of LOD. Considering the sparse energy distribution of speech signals in the time-frequency domain, the Dominant LOD may arise in some short time subbands even though it appears to be Non-dominant LOD in its fullband. Therefore the proposed LOD-based ICA is extended to the frequency subbands for more opportunities to attain such Dominant LOD type.; The effectiveness of the proposed short time LOD-based ICA is validated by applying it to a speaker-moving model and a mixing system with abrupt changes, which approaches the practical applications better since the mixing system is not always constant as in standard ICA model. We have also explored the separation task with noise-contaminated speech signals. This suggests us that: other than the long time analysis, the short time analysis may provide an alternative means with extra information for separation when the independence information is impaired and subsequently fails to yield the desirable separation performance.