Sunday, August 08, 2004

IBM Research: Hindi Audio Visual Speech Recognition

"At the IBM India Research Lab (IRL), we are extending the IBM ViaVoiceTM recognition technology to build a speech recognition system for the Hindi language. We have identified a phone set consisting of 61 phonemes for Hindi. A mapping from the English phone set to Hindi phone set was used to bootstrap the English acoustic model to build the initial phone models for HindiP2. Using these models, we have built an acoustic model for Hindi by training it on a large sample of Hindi acoustic data. We have the setup to collect and transcribe speech samples that are used to train the acoustic models.

"We have built a statistical language model for Hindi, which captures the language statistics from a huge amount of text data. Currently we are working on increasing the accuracy of our Hindi speech recognition system by enhancing the Acoustic and Language models. The recognition system performs well over a trained context and we have a demo of a large vocabulary continuous speech Hindi recognition system at the Lab. Our goal is to cover more Indian languages and then to build a multilingual speech recognizor for the Indian languages based on a multilingual phone set."