Within- and between-speaker acoustic variability: Spontaneous versus read speech

Abstract

Using principal component analysis (PCA), our previous study [JASA 145(pt. 2), 1930, (2019)] of read sentences found surprisingly similar acoustic voice spaces for groups of female and male talkers and for the individuals within groups. Formant frequencies and the balance between higher harmonic amplitudes and inharmonic energy in the voice accounted for the most acoustic variance within and across talkers, but many further details varied idiosyncratically for individual talkers. In this study, we replicated this finding using a set of recorded phone conversations from 99/100 original speakers (49 F), hypothesizing that the same measures would characterize both individual and population acoustic spaces, despite greater acoustic variability for spontaneous utterances. F0, formant frequencies, spectral noise, source spectral shape, and their variability were measured every 5 ms from vowels and approximants. Individual and group PCAs revealed that the acoustic voice spaces derived from spontaneous speech are highly similar to those spaces previously identified based on read speech. One significant difference between the two speaking styles was that unlike read speech, variability in F0 emerged as one of the variables that accounted for significant acoustic variability in spontaneous speech. Implications for voice learning, recognition, and discrimination will be discussed. [Work supported by NIH/NSF.]