Abstract
Six low-bandwidth measures were used in three types of programs for the automatic recognition of spoken digits. The measures were chosen to be closely related to articulatory rather than to acoustic properties of speech. The first program, without any learning feature, asked specific questions about the values of the six measures; its accuracy ranged from 64% to 97% correct. Two speaker/specific programs, which learned with a sample of two utterances per digit, yielded accuracies averaging 97% when tested on new utterances from the same talker. Performance fell to 88% and 94% when learning was carried out on a pool of four speakers, and to 78% and 86% when a 3-speaker pool provided the learning for classification of a fourth speaker's utterances. It is suggested that such “nonacoustic” measures can be of substantial value in more-general speech-recognition procedures.
Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
