In this paper, we present a unified approach for hidden markov model based multilingual speech recognition. The proposed approach could be used across acoustically similar as well as diverse languages. We use an automatic phone mapping algorithm to map phones across languages and reduce the effective number of phones in the multililingual acoustic model. We experimentally verify the effectiveness of the approach using two acoustically similar languages, Tamil and Hindi and also American English which is very different from the other two languages acoustically. The experimental results are very encouraging and demonstrate the effectiveness of the approach in building a universal multilingual speech recognition system.
Dr. Santhosh Kumar C., Mohandas, V. P., and Li, H., “Multilingual Speech Recognition: A Unified Approach”, in InterSpeech, 2005.