HMM adaptation using statistical linear approximation for robust automatic speech recognition

The lack of noise robustness is one of the main drawbacks of an Automatic Speech Recognition (ASR) system. A well trained ASR system can achieve high recognition rate on quiet laboratory conditions, but perform poorly in real life environments. In this paper we will present a noise robustness method which uses the clean speech Hidden Markov Models (HMM) and noise statistics, to create an approximation of the degraded speech HMM using the Statistical Linear Approximation (SLA). Experiments using the proposed methods had shown up to 87.7% word error rate improvement.