Sinhala Speech Recognition with Speaker Adaptation

Abstract

This research describes a successful way of building a Sinhala speech recognizer using an already built speaker dependant continuous Automatic Speech Recognizer (ASR) with speaker adaptation techniques. Sinhala which is spoken everywhere in Sri Lanka has different variations in different parts of the country. Though Sri Lanka is a small island there are many difficulties in building a successful speaker independent automatic speech recognition solution for general use. Therefore in this research we describe a solution for adapting a speaker dependant Sinhala speech recognizer into another speaker s voice. The proposed solution consists of building a successful speaker dependant speech recognizer using the free and open source toolkit CMUSphinx and adapting it to recognize speech of another speaker. The original speaker dependent Sinhala speech recognizer was trained using 200 sentences consisting nearly 1000 distinct words with an accuracy of 71.86% on words in the vocabulary. This recognizer was then adapted individually to male and female speakers using just 30 training sentences of each. While their accuracies against the original model were below 35%, the adapted models were 75% or more accurate based on the word error rate metric.