Continuous speech segmentation determined by blind source separation
Harold H Szu, Charles Chia-chuen Hsu, Dahong Xie · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1998
One of the problems of 5 percent error rate encountered in continuous speech recognition is partly due to the difficulty in the identification of a mixed up to two phonemes in a close concatenation. For instance, one speaks of 'Let's go' instead of 'Let us go'. There are two kinds of speech segmentations: the linguistic segmentation and the acoustic segmentation. The linguistic segmentation relies on a combination of acoustic, lexical, semantic, and statistical knowledge sources, which has been studied. Daily spoken conversations are usually abbreviated for speakers' convenience. The acoustic segmentation is to separate the mixed sounds such as /ts/ into /t/ and /s/ for automatically finding linguistic units. Adaptive wavelet transform (AWT) developed by Szu is a linear superposition of banks of constant-Q zero-mean mother wavelets implemented by an ANN called a 'wavenet'. Each neuron is represented by a daughter wavelet, which can be an affine scale change of identical or different method wavelet for a continuous AWT. AWT was designed for the cocktail party effect and to solve the acoustic segmentation of phonemes using a supervised learning ANN architecture. In this paper, we reviewed AWT from Independent Component Analysis viewpoint, and then applied blind source separation to the acoustic de-mixing and segmentation.