Pilot experiment to show that modifications in pitch, pitch range, and formants of speakers influence the perceptions of voice likability.
By Laura Fernandez Gallardo and Benjamin Weiss. Thanks to Elvira Ibragimova for Praat scripts for speech manipulation.
Selecting speakers from the NSC corpus to be manipulated (originally positively and negatively perceived): by Benjamin Weiss. Benjamin also manipulated "manually" some files which had artefacts when manipulated with Praat: m041_bucharest (pitch), m041_bucharest (pitch and pitch range), m090_riyadh (pitch).
Praat scripts for speech manipulation: written by Elvira Ibragimova.
Listening tests have been conducted with 7 participants (5m, 2f) who rated perceived warmth (on the scale "cold----hearty" - referring to speakers' warmth) of manipulated speech. They were aged 27 years on average (stdev = 2.3, range: 23-30). Their mother tongue was different: Croatian, Bangla, Polish, Nepalese, Spanish, German, Dutch.
The speech stimuli consisted of 8 speakers and 5 manipulation conditions. These involved the variation of: mean pitch, pitch range, combination of mean pitch and pitch range, combination of mean pitch and formant shift.
A 3-way repeated measures ANOVA tested the effects of the independent variables on perceived speakers' warmth:
- speech manipulation condition
- speaker gender
- speaker original WAAT (warmth-attractiveness)
There is a significant effect of 'condition' on 'herzlich' (p<0.01), that is, the manipulation of speech affects the perceived warmth from speakers. No effect has been found for speaker gender or for speakers' original warmth (positive or negative).
However, our initial hypothesis:
- lowering mean F0 and formant frequencies and increasing F0 range of low-WAAT speakers will result in higher perceptions of warmth.
- increasing mean F0 and formant frequencies and lowering F0 range of high-WAAT speakers will result in lower perceptions of warmth.
are generally not met.
In view of these results, our hypothesis that high-WAAT speakers would be perceived as less warm with an increase of F0 mean and decrease of F0 range is not met. Lowering F0 mean did not contribute of the perception of low-WAAT speakers as warmer. Only the increase of F0 range of low-WAAT speakers caused these voices to be perceived as warmer.
Hence, a different listening test is proposed for future work, in which voices will be presented in a paired-comparison setup. Original and manipulated speech segments corresponding to the same speaker will be presented, and the listeners will be asked to select that segment presenting higher warmth. This paradigm is expected to offer more conclusive insights and meet our initial hypothesis.