On the use of F0 variations in automatic speech recognition

Abstract

This study deals with the feasibility of using microprosodic pitch variations in CVC words in automatic speech recognition. It has been demonstrated in the past [K. Silverman, Proc. Inst. Acoust. 6, 445–452 (1984)] that the fundamental frequency (F0) contour at a post‐release voicing onset of CVC words can serve as a cue in distinguishing between voiced and voiceless consonants. The word pitch contours in the CVC speech database obtained from MIT were examined. The word‐initial consonants in this study involved stops /p, t, k, b, d, g/ and affricates /ts, dž/, whereas the vowels included /i, ᴜ, ae/. The algorithm that was used in pitch measurement is an average magnitude difference function (AMDF). With this method, pitch is estimated by running a half‐wave rectified LPC residual of the speech waveform through the AMDF. Pitch measurements confirm that, for a majority of CVC words in this database, fundamental frequency in the initial glottal cycles of a post‐release vowel is higher for voiceless consonants...