Effects of speaker accent on the performance of a speaker-independent, isolated-word recognizer
V. Gupta, Paul G. Mermelstein · The Journal of the Acoustical Society of America · 1982
We report on an investigation of comparative performance of a speaker-independent, limited-vocabulary, isolated-word recognizer when trained and tested on (1) native English–Canadian speakers, and (2) French-Canadian speakers of English. Randomized 20-word vocabularies were recorded from 20 male and 20 female speakers of each group over individually dialed and locally switched telephone connections. Recognition performance with two clustering techniques was evaluated and a complete-link clustering algorithm is found to yield consistently better results. Although a recognition rate of 97% is attained when using ten templates/word for the native speaker group, the same rate drops to 93% for the foreign-accented speakers. For the linguistically homogeneous group, no significant increase in recognition rate is noted beyond four templates/word, whereas for the nonhomogeneous group, performance is still improving at ten templates/word. Additionally, the variability in recognition rate for individual speakers is significantly larger for the French Canadians. Although the above recognition rates may be acceptable for many applications, for even better performance it appears necessary to build into recognizers additional speech-specific knowledge such as differences in acceptable acoustic variation of phonetic segments.