Phonetic recognition in a segment-based HMM

The author describes a segment-based HMM (hidden Markov model) recognizer and presents phonetic recognition results achieved with the system. As opposed to a conventional frame-based HMM, measurements in such a system are made on variable-duration segments. The key experimental result is that inclusion of measurements made beyond segment boundaries improves phonetic recognition performance significantly. On a set of nine male test speakers from the VOYAGER corpus, the system obtained a phonetic recognition accuracy of 59% (95% confidence interval of 53-65%) on a 39-class phonetic recognition task. Although little attempt was made to optimize system parameters, this result is competitive with existing systems of comparable complexity.<>