Developing Tigrinya Speech Recognizer Using Amharic and Tigrinya Data
No Thumbnail Available
Date
2015-03
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
This study has introduced the design of a Hidden Markov Model based LVCSR system in a new target language based on a different source language and without the need of a large speech databases on the target language. The Tigrinya LVCSR was developed using an Amharic Corpus consisting of 10,850 sentences and a limited Tigrinya data containing 600 sentences to train the acoustic models. The study was conducted based on the knowledge based approach taking the assumption that the articulatory representations of phonemes are similar across the Tigrinya and Amharic languages with the exception of 2 phonemes unique to Tigrinya and using all the phonemes as acoustic units.
A total of six experiments were performed using different parameters each one done in an effort of increasing the performance of the recognizer. Out of the five experiments, the best result obtained with the experiment that is done by training the seed model with the 10,850 Amharic sentences up to the 8th iteration and using the 600 Tigrinya sentence starting from the 8th iteration of the training process. The experimental result showed percentage of correctly recognized words of 88.33% with an accuracy of 73.43 %.
The baseline Tigrinya recognizer which was trained using only the 600 Tigrinya data resulted in correctly recognized words of 80.80% and an accuracy of 67.39 % on the tri-phone model with 12 Gaussian mixtures. Comparing this result with the best result obtained in the experiment showed that an increase of about 8% was achieved in terms of correctly recognized words and of about 6% in terms of accuracy. This has proven that the use of Amharic data with limited Tigrinya data for training a Tigrinya recognizer does result in significant performance increase and that it is a promising future research direction given that different methods are applied to further achieve better results. As this is the first attempt other phone mapping techniques and approaches such as the data driven approach can also be tried for performance improvement purpose.
Description
Keywords
Speech Recognizer