Etienne Denoual
Noise robust sub-band ASR using a HMM/BN framework
Abstract:This document will present the research work I conducted in ATR for five months
from April to August 2003. Many approaches have been proposed in the past to deal
with noise in robust speech recognition, mainly feature-based and model-based.
Here we consider the use of a hybrid HMM/BN framework to increase ASR
robustness, by computing the posterior likelihood of noise in each frame's sub bands,
and then marginalizing corrupted channels.