Login
Talks & Seminars
Title: Discriminative Adaptive Training and Adaptation for Speech Recognition
Dr. Chandrakant K. Raut,
Date & Time: December 5, 2011 11:00
Venue: Conference Room, 01st floor, C Block, Dept. of CSE, Kanwal Rekhi Building
Abstract:
In many automatic speech recognition systems, adaptation and adaptive training plays an important role to deal with varying speakers or acoustic environment. Though the acoustic models are usually trained using a discriminative criterion such as minimum phone error in the state-of-the-art systems, adaptation transforms are still usually obtained through maximum-likelihood (ML) estimation. This is because discriminative transforms are biased towards the supervision hypothesis, and are very sensitive to any errors in it. Therefore, they give only a little gain, if any, for unsupervised adaptation. Moreover, there may be only a small amount of adaptation data, specially in case of instantaneous adaptation, which may not yield robust estimates of the transforms. In this talk, these hypothesis bias and data sparsity problems of discriminative adaptation will be addressed through the use of criterion mapping functions and a Bayesian framework. An adaptive training scheme based on discriminative mapping transforms (DMT) will be first described. A DMT is a speaker-independent transform to map ML transforms into discriminative ones. It does not directly depend upon the test supervision hypothesis, and thus is not sensitive to any errors in it. This will be followed by a description of a Bayesian approach to adaptation that treats transforms as random variables and uses prior distributions for them. Finally, the Bayesian approach will be combined with the DMT to yield a discriminative instantaneous adaptation framework.
Speaker Profile:
C. K. Raut received PhD in Information Engineering from Cambridge University, Machine Intelligence Laboratory (MIL). He has an experience of working as a scientist in the area of speech and language processing at Raytheon BBN Technologies, Massachusetts, USA. He has been involved in Defense Advanced Research Project Agency's (DARPA's) Global Autonomous Language Exploitation (GALE) and Robust Automatic Transcription of Speech (RATS) projects. He has worked in the area of robust speech recognition under noise and reverberation, acoustic modelling, adaptive training, and model adaptation. He has been a reviewer in several conference and journal publications in the area of speech and language processing. He is also a recipient of the NEA Young Engineer Award.
List of Talks

Webmail

Username:
Password:
Faculty CSE IT
Forgot Password
    [+] Sitemap     Feedback