Longer-length acoustic units for continuous speech recognition

Recent research on the TIMIT database suggests that longer-length acoustic units are better suited for modelling pronunciation variation and long-term temporal dependencies in speech than traditional phoneme-length units, yielding substantial improvements in recognition accuracy [9]. In this paper, we investigate whether similar improvements can be gained on another database, viz. excerpts from novels in a Dutch library for the blind. We use a hierarchical method that employs a mixture of word-, syllable- and phoneme-length units. Our results show that the approach does increase the word accuracy, but to a lesser extent than expected. The paper discusses possible explanations for the finding.