Recommendations for advancing speech recognition

Abstract

Surveying the opinions of over 100 workers in speech recognition, we found widespread agreement about the gaps in current technology and the needs for future research and development. Improved phonetic segmentation and labeling is considered a top priority need in future studies. Also frequently mentioned are phonological rules (covering coarticulation effects, phone substitutions, deletions, and insertions, and word boundary effects); and prosodic aids to providing more efficient phonemic analysis and syntactic parsing. Work should continue on strongly constraining hypothesized word sequences with syntactic and task‐dictated constraints, while permitting naturalness of expression. The practical needs for continuous speech are not well documented, and the wide gap between available isolated word recognizers and experimental limited speech understanding systems needs to be bridged by work on several systems of varied complexities and capabilities. Performance evaluation should be given special attention. We recommend the establishment of speech science centers with speech and linguistics expertise, powerful computer facilities, and a mechanism for visiting researchers to use such facilities to advance their work and refine the resident recognition systems. [Work supported under Office of Naval Research Contract Number N00014‐77‐C‐0570.]