Acoustic analysis of vocal effort and speaking style

Abstract

“The singer’s formant” is used by singers during voice projection. While a similar “actor’s formant” has been found (Leino, 1993), other work finds greater source effects in projected voices (e.g., Master et al., 2012). It remains unclear whether differences in source and/or filter remain among non-actors/singers in varying speaking styles. Recordings from the SRI-FRTIV corpus (Shriberg et al., 2008) were used to examine the effects of vocal effort and speech style on vocal characteristics. 34 speakers of American English spoke in up to three levels of vocal effort in (1) Interview, (2) Conversation, (3) Reading, and (4) Oration tasks. The following measures were measured by LTAS based on the mid-50% of the recording’s duration: (a) CPPS; (b) L1-L0; (c) alpha ratio; (d) mean frequency of the lowest peak; (e) the level difference between strongest peaks at 3–4 kHz and 0–1 kHz (“the actor’s formant”); the level difference between the first formant region (300–800 Hz) and the (f) 1–2 kHz, (g) 2–3 kHz, (h) 3–4 kHz, and (i) 4–5kHz range. Preliminary results using PCA find that general measures of spectral slope (PC1: h, i, c, g, b) better capture differences in vocal effort and speech style than more fine-grained source/filter measures (PC2: e, f, a, b, d).