Contribution of vocal tract and glottal source spectral cues in the generation of happy and aggressive [a] vowels

Abstract

At present, three-dimensional (3D) acoustic models allow for the numerical simulation of vowels, diphthongs and some vowel-consonant-vowel sequences using realistic vocal tract geometries. While research is being done to generate more phonemes and short utterances, some attempts have been made to incorporate expressiveness into the 3D numerical simulation of isolated vowels. However they are very preliminary and still far from the generation of expressive utterances. To move towards this goal, this work analyses the contribution of vocal tract (VT) and glottal source spectral (GSS) cues to the production of happy and aggressive vowels with respect to neutral vowels. After parameterising with the GlottDNN vocoder the paired neutral-expressive utterances from a Spanish database, neutral utterances are transplanted with the target expressive prosody as baseline, and subsequently resynthesised considering also the GSS and/or VT from their expressive pairs. Objective and subjective evaluations show that, both GSS and VT have a statistically significant contribution to convey the tense voice target emotions. VT prevails over GSS specially for aggressive. Best results are achieved when considering both GSS and VT, which compared to the baseline permits an increase in the perceived emotional intensity of 55.3% for happy and 62.8% for aggressive utterances.