First International Conference on Spoken Language Processing (ICSLP 90)
This paper describes performance evaluation in speech recognition system which uses transition probability between linguistic units. The lower limit of word recognition score is predicted based on phoneme recognition score and number of word pairs with short distance in a vocabulary defined by linguistic information. But it is difficult to calculate it when transition probability is used as linguistic information. We propose new algorithm to calculate it when bigram or trigram of linguistic units is used. Using this algorithm, we carry out performance prediction in speech recognition which uses bigram or trigram. Recognition score for word with 5 phonemes is more than 26% using bigram, more than 71% using trigram and more than 95% using a dictionary when phoneme recognition score is 90%, where bigram and trigram of phonemes are estimated from the 5,317 Japanese popular words. Recognition score of sentence composed of 11 words is more than 4.3% using bigram, on the other hand, more than 67% using trigram, when word recognition score is 80%, where bigram and trigram are estimated from 136 sentences represented with 18 kinds of speech.
Bibliographic reference. Otsuki, Takashi / Makino, Shozo / Sone, Toshio / Kido, Ken'iti (1990): "Performance evaluation in speech recognition system using transition probability between linguistic units", In ICSLP-1990, 1213-1216.