Differences of speech rate, interphoneme distance and likelihood caused by speaking style, their relationship, and recognition performance

Abstract

AbstractWe have been developing a speech recognition system as part of a spoken dialogue system. There is a clear limit of recognition performance for dialogue speech using acoustic models built with a read speech database, mainly because various acoustic and linguistic phenomena, which reflect the characteristics of spontaneous speech, occur in dialogue speech. We made acoustic models of syllable unit using simulated dialogue speech or read speech database. We then analyzed the acoustic properties of the dialogue speech in terms of the distance between models and the distribution of duration and compared them with read speech. We found that the speech rate of dialogue speech was faster on average and had a greater variance of speech rate than read speech. We also found that the distance between vowels of the dialogue speech tended to be smaller than that of read speech, that is, the vowels tend to be confused in dialogue speech because of closer boundaries to one another. We also compared recognition performance using acoustic models trained with each database and found that dialogue speech had a low correct syllable rate. Dialogue speech was more prone to insertion and deletion errors and the influence of distance between vowels and the variance of speech rate was shown clearly in the results. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(7): 50–60, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.1141