EXTRACTING SPEECH-RATE VALUES FROM A REAL-SPEECH DATABASE.

Abstract

An examination of the articulatory and acoustic correlates of duration change using rules based on work by D. H. Klatt (1979) confirms the influence of such known effects as stress, phonetic context, intrinsic vowel duration, part of speech, and position in utterance and provides a key to quantifying and isolating the effects of speech rate through measurement of the smoothed error extracted from the difference between predicted and observed durations. A database of 5000 syllables from the Spoken English Corpus, measured for duration and tagged for features of stress, part-of-speech, and structural context, was prepared. Syllables were measured with the SAY speech analyzer using interactive graphic display and auditory replay of the waveform. A tagged phonemic transcription was used as input to modified Klatt rules, and the resulting duration predictions compared against measurements in the database. Preliminary results show that the Klatt rules are flexible enough to provide the desired quantification.