A new approach to develop a syllable based, continuous Amharic speech recognizer

All of the previous syllable based Automatic Speech Recognizers (ASRs) for the Amharic language are built by training a separate acoustic model for each of the 196 distinctly pronounced Consonant-Vowel (CV) syllable. In this paper, we will demonstrate that a smaller number of acoustic models are sufficient to build a syllable based, speaker independent, continuous, Amharic ASR. It is built for weather forecast and business report applications using the UASR (Unified Approach to Speech Synthesis and Recognition) Tool kit. A new speech corpus, which is of more than 35 hours duration, is used for training. It is a collection of corpora recorded in three different environments in order to make the recognizer less sensitive to recording environment and microphone changes. The grammar is finite state transducer based and the lexical model consists of thousands of words. Though acoustic models for only 93 syllables are trained, a recognition accuracy of 93.26% is achieved on a test set that has 4,000 words collected from 10 speakers.