Efficient Automatic Speech Recognition

D. O'Shaughnessy (Canada)


Automatic speech recognition, analysis, features, spec trum.


Automatic recognition of certain speech (e.g., tele phone numbers) is very feasible today with very good ac curacy, even when using telephone lines and serving a large population. However, even such simple recognition tasks suffer decreased performance in adverse conditions, e.g., significant background noise or fading on portable tele phone channels. If we further impose significant limita tions on the computing resources for the recognition task, then robust and efficient speech recognition is still a signif icant challenge, even for small vocabularies (e.g., digits). The amount of computer resources needed for good recog nition accuracy was investigated. Rather than use a tradi tional hidden Markov model approach with cepstral analy sis, which can be computationally intensive and often does not work well under adverse acoustic conditions, a simpler spectral analysis was used, combined with a segmental ap proach. High recognition accuracy can be maintained de spite a large decrease in both memory and computation.

Important Links:

Go Back