Speech Analysis and Recognition Synchronised by One-Quasiperiodical Segmentation

It is shown that the best ASR results are attained when a pre-processing is carried out synchronically with pitch. Specifically, an analysis step has to be equal to the current one-quasiperiod duration and current analysis intervals have to consist of an entire number of quasiperiods with total 45-60 ms duration. Quasiperiodicity and non-qusiperiodicity models and measures as well as their applications for the optimal segmentation of speech signals into one-quasiperiods are given and discussed. Then the ways to embed these pre-processing results into the recognition procedure are described.

• Type: Conference
• Year: 1999
• Where: TSD
• Authors: Taras K. Vintsiuk, Mykola M. Sazhok
• Keywords: Current Analysis Intervals, Current One-quasiperiod Duration, Signal Processing, Total 45-60 Ms, TSD 1999