SPEAKER INDEPENDENT TELEPHONE SPEECH RECOGNITION USING SUBWORD UNITS

Abstract

An inventory of HMMs has been generated for the development of speaker and vocabulary independent speech recognition applications in telephony.This has been achieved by applying a modelling technique which uses subword units (biphonicles) to a fairly large telephone speech corpus.The best average recognition error rates for ve different task vocabularies were 6.6% and 6.8% for male and female speakers respectively.6.