Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Prediction of Word Confusabilities for Speech Recognition

David B. Roe, Michael D. Riley

AT&T Bell Laboratories, Murray Hill, NJ, USA

Words which are similar in pronunciation cause errors by speech recognizers. In an application of speech recognition, the vocabulary should be chosen so as to avoid similar sounding words or phrases. Phonetically similar words (such as "wait" and "eight") or short words ("on" and "off) may be confused by the speech recognizer, with undesirable consequences. We have developed a software tool, word_confuse, that detects confusable words. The confusability between pairs of words is calculated from two sources of information: the phonetic pronunciation of words as determined by the AT&T text-to-speech synthesizer, and the phonetic confusions exhibited by an AT&T phone-based speech recognizer. The calculation of confusability is based on searches through a finite state network that represents probabilistically the phonetic pronunciation of words. The metric of similarity is based on the Bhattacharyya distance. Word_confuse can be used to detect and eliminate confusable words from the vocabulary used in speech recognition applications.

Full Paper

Bibliographic reference.  Roe, David B. / Riley, Michael D. (1994): "Prediction of word confusabilities for speech recognition", In ICSLP-1994, 227-230.