Acoustic training system for speaker independent continuous Arabic speech recognition system

Abstract

The paper presents an acoustic training system for building acoustic models for a medium vocabulary speaker independent continuous speech recognition system. A speech database is constructed to train the acoustic models. The acoustic models are constructed, and trained. A test set database is constructed to test the accuracy of the acoustic models. Also 4 language models of two main types: bigram and context free grammar, were built and used in tests. Our results show 5.26% and 2.72% word error rates for 1340 and 306 words bigram based language models, respectively. Our results show also 0.19% and 0.99% word error rates for 1340 and 306 words context free grammar based language models, respectively.