Segmentation of continuous speech by using multidimensional scaling techniques

Continuous speech is digitized at the rate of 20480 Hz. Power spectra are taken on 1024 blocks shifted from 256 to 256 samples. These spectra are divided into 25 channels chosen to discriminate at best peaks and valleys. For each spectrum i, and each channel j, the sum Pij of the components is computed. A multidimensional scaling analysis is done on the matrix Pij. This gives a 7-dimension space in which variations of the spectra versus time are represented by a moving spot. The main result is that a transition between two spoken sounds induces a significant variation on, at least one, and generally several axes. This can be used for segmenting and recognizing the continuous speech at the acoustical level. The first attempts have given modest results, but great improvements are expected soon.