Steady part recognition of continuous speech for acoustic-phonetic decoding

We present in this paper an automatic segmentation and labelling module for the steady parts of continuous speech. The segmentation stage splits the signal into five classes, and the labelling stage uses vector quantization techniques as well as an original decision rule mixing the Bayes and K Nearest Neighbor criteria. The resulting (single speaker) recognition scores are very encouraging. This project is supported by grants from ANRT(*) and TELIC-ALCATEL, and realized at ENST.