Robust Speech Recognition by Combining a Robust Feature Extraction with an Adaptation of HMMs

Abstract:
A method is presented to extract robust features from a noisy speech signal with the intention to improve the performance of an automatic speech recognition system. The processing is based on an adaptive filtering of the short-term spectra where the frequency response of the filter is smoothed with a cepstro-temporal approach. It turns out that the recognition performance is comparable with the performance that can be achieved with a robust feature extraction scheme standardized by ETSI. Looking at the case of a hands-free speech input in a noisy and reverberant environment the recognition rates can be improved further by additionally adapting the HMMs to the acoustic conditions