Using multi-level segmentation coefficients to improve HMM speech recognition

This paper presents a new kind of acoustic features for HMM speech recognition. These features try to capture phone-specific segmentation information using multiple temporal resolutions. Experiments show that word accuracy can be improved by 7% when combining these features with traditional mel-cepstral coefficients in a speaker-independent word recogniser. This improvement is mostly due to a reduced number of insertion and deletion errors.