A Markov model acoustic phonetic component for automatic speech recognition

Abstract

A Markov-model acoustic-phonetic component is constructed for the synthesis of standard acoustic representations of connected speech. The primary building blocks are phones with Markov models structured so that phone length, spectral power and fundamental frequency are parametrically controlled. The model generates acoustic parameter outputs at 10-ms time steps. The acoustic-phonetic component permits matching between actual acoustic data and internally modeled acoustic data, and can be employed in various ways - to automatically label speech, as a phone decoder to obtain estimated phone strings, and in speech recognizers which match at the acoustic level.