Incorporating voice onset time to improve letter recognition accuracies

We consider the possibility of incorporating distinctive features into a statistically based speech recognizer. We develop a two pass strategy for recognition with a standard HMM based first pass followed by a second pass that performs an alternative analysis to extract class-specific features. For the voiced/voiceless distinction on stops for an alphabet recognition task, we show that a linguistically motivated acoustic feature exists (the VOT), provides superior separability to standard spectral measures, and can be automatically extracted from the signal to reduce error rates by 48.7% over state of the art HMM systems