Incorporating the concept of the syllable into automatic speech recognition may improve recognition accuracy by helping to integrate information over syllable-length time spans. Evidence from psychoacoustics and phonology suggests that humans use the syllable as a basic perceptual unit in speech processing. Nonetheless, the explicit use of such long-time-span units is comparatively unusual in modern automatic speech recognition systems for English.

The work to be described in this talk explored the utility of information collected over syllable-based time-scales. The first approach involved integrating syllable segmentation information into the speech recognition process. The addition of acoustically-estimated syllabic onsets resulted in a 10% relative reduction in word-error rate. The second approach began with developing four speech recognition systems based on long-time-span features and units, including modulation spectrogram features. Analysis suggested the strategy of combining, which led to the implementation of methods that merged the outputs of syllable-based recognition systems with the phone-oriented baseline system at the frame level, the syllable level and the whole-utterance level. These combined systems exhibited relative improvements of 20-40% compared to the baseline system for clean and reverberant speech test cases.