10th Speech in Noise Workshop, 11-12 January 2018, Glasgow

Acoustic analyses of vowel variation for the investigation of perceptual adaptation to speaker properties in channel-vocoded speech: preliminary data

Olivier Crouzet(a)
LLING - Université de Nantes / CNRS, France

Etienne Gaudrain(b)
CNRS, University of Lyon, Lyon Neuroscience Research Center, France

Deniz Başkent(b)
University of Groningen, University Medical Center Groningen, Netherlands

(a) Presenting
(b) Attending

Perceptual analysis of speech requires processing two complementary dimensions: on one side, various acoustic parameters provide access to information concerning voice properties (vocal characteristics, speaker identity, gender, emotional state...) while, on the other side, listeners process sounds into phonological categories (phonemic classes / distinctive features). These two dimensions interact strongly in speech. For instance, when comparing speakers with different vocal tract lengths (VTL), a shorter VTL would be associated with higher "formant frequencies" overall. When comparing different vowels, a single speaker contrasting the vowels /o/~/u/ would lower both the first and second formants by changing the shape of one's own vocal tract. These interactions seem to have a relatively low impact on everyday communication in normal-hearing listeners: it is assumed that some perceptual adaptive mechanisms take place that contribute to the listeners' ability to deal with these interactions. However, these phenomena may constitute a source for the low speech recognition performance that some cochlear-implanted deaf listeners experience in multi-speaker situations. In order to investigate this issue, we are setting-up a series of experiments in order to investigate the influence that adaptive mechanisms to changes in voice properties (e.g. vocal tract length) may exert on phonological classification (e.g. vowel identification) when listening is degraded using channel-vocoded speech.

We plan to replicate a classical experiment (Ladefoged & Broadbent, 1957; see also Sjerps, McQueen & Mitterer, 2013) in order to measure the level of vowel normalization that speakers may reach when processing vocoded speech in contrast to natural speech. In order to address these issues, formant measurements that were issued from a database of natural recordings of Dutch speakers were investigated in order to assess both formant frequency central tendencies and variation. These measurements provide access to information concerning between-vowel average distance and within-vowel individual variation with respect to (1st, 2nd and 3rd) formant frequencies. They also provide us with observations concerning the direction of change from one vowel category to the other for each formant and their relation to the expected impact of differences in vocal-tract length. From these data, a set of 15 Dutch monosyllabic word-pairs were selected and recordings by 3 Dutch speakers were collected. For each vowel contrast (word pairs), an acoustic continuum was generated on the basis of the actual formant frequencies. These results will be described along with a preliminary crowd-sourcing experiment that will provide measurements of psychometric identification curves for each vowel continuum.

Last modified 2017-11-17 15:56:08