IRL ready with Hindi speech recognition software

NEW DELHI: The IBM Research Lab (IRL) at IIT Delhi is ready with a prototype for Hindi speech recognition software. However, it is yet to be made available as a product as many finer features to the software are yet to be incorporated. One has to train the software for about half an hour to customize the pronunciation of individual speakers. The center is currently working on increasing the accuracy of Hindi speech recognition system by enhancing the acoustic and language models.

Essentially, what IRL has done is to extend IBM ViaVoice voice recognition technology to build a speech recognition system for Hindi. It identifies a phone set consisting of 61 phonemes for Hindi. Then it uses a mapping from the English phone set to Hindi phone set to bootstrap the English acoustic model to build the initial phone models for Hindi. Using these models, the Center then built an acoustic model for Hindi by training it on a large sample of Hindi acoustic data.

One of the major focus areas of IRL has been speech recognition software. However, a huge challenge for the Speech Group at IRL has been the design and implementation of computer systems that can cover the whole range of human like interaction by using faces and voices. The group aims to build speech recognition systems for Indian languages and improve the existing speech recognition techniques for them to be more useful in the real world by making them more robust to ambient noise.



The Center is said to be currently in talks with a number of organizations which would help to take the technology to a greater number of people. Manoj Kumar, Director at the Center said that the organization was talking to Media Labs Asia for deploying the software in rural India.

Speech recognition technologies can be a vital step in bringing the Information Technology world to the Indian masses. Using such a convenient means of rendering information to and from the machine would mean that the end user need not be computer literate and still use the power that the IT industry brings to the society.






Besides Indian speech recognition, IRL is also working on telephony speech recognition. Speech recognition over the telephone line is a more challenging task owing to the 8KHz band limited speech signal and the noise introduced in the telephone channel. Also, the speaker is more spontaneous in the conversations. "However, by training a recognition system on the context in which it is to be used, it is possible to increase the recognition accuracy of the system for that particular domain," informed Kumar. For example, telephony speech recognition can be used in automating and improving telephone banking, directory assistance, inquiry systems, he added.