Efficient automatic speech recognition

Automatic recognition of certain speech (e.g., telephone numbers) is very feasible today with very good accuracy, 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 telephone channels. If we further impose significant limitations on the computing resources for the recognition task, then robust and efficient speech recognition is still a significant challenge, even for small vocabularies (e.g., digits). The amount of computer resources needed for good recognition accuracy was investigated. Rather than use a traditional hidden Markov model approach with cepstral analysis, 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 approach. High recognition accuracy can be maintained despite a large decrease in both memory and computation.