Feature enhancement for robust speech recognition

The results of investigations into some aspects of robust speech recognition are reported in this thesis. Included in the topics that have been studied are feature extraction, training and decoding procedures, speech feature enhancement and model adaptation. In an automatic speech recognition (ASR) system, feature extraction is critical to determining system performance. The most commonly used feature vectors for ASR are those based on the Mel Frequency Cepstral Coefficients (MFCC). However, it is well known that under noisy conditions, the performance of MFCC-based speech feature vectors degrades significantly. There have been many other robust features proposed in recent years and one that is derived from phase autocorrelation (PAC) was investigated.MASTER OF ENGINEERING (EEE