Perceptual Learning of Natural and Sinewave Voices
Author: Sonya M. Sheffert, David B. Pisoni, Jennifer M. Fellowes, and Robert E. Romez
Abstract:
This report describes the results of a perceptual training study that was
designed to explore how listeners learn to categorize novel voices and how
knowledge of a familiar voice generalizes to novel utterances. The speech
samples from which the listeners learned to identify individual were of
two kinds: Naturally produced English sentences and sinewave replicas of
these sentences. The sinewave items were nonspeech tonal patterns that preserved
coarse-grained properties of the talker's vocal tract transfer function
while eliminating traditional cues to voice quality. Listeners were trained
over several days to identify by name ten talkers from sentence length sinewave
or natural speech utterances. Knowledge about the talker's voice was then
assessed using two generalization tests in which listeners heard a novel
set of sentences and were required to identify the speaker. In one generalization
test, the sentences were sinewave replicas whereas in the other generalization
test, the sentences were naturally produced. The results showed that perceptual
learning of a talker's voice can occur even when specific acoustic products
of vocal articulation are eliminated from the signal. The data also showed
that speaker-specific knowledge acquired during this perceptual training
task generalized to novel natural and novel sinewave sentences. Variability
in the degree of perceptual learning affected generalization of speaker
knowledge. The results of this study show that listeners can learn about
a talker's voice form highly impoverished acoustic signals when the products
of vocal articulation are eliminated, and that this knowledge generalizes
to novel utterances produced by these same talkers.