A Hierarchical Bayesian Model of Multi-level Phonetic Imitation

It is well known that listeners adapt, in some sense, to speech that they have recently heard. Words spoken in recently heard voices or accents are recognized more quickly and accurately (Mullennix et al. 1989; Goldinger 1996; Nygaard & Pisoni 1998; Maye et al. 2003; Kraljic and Samuel 2006, 2007; Smith 2007; see Nygaard 2008 for a review). And listeners can become attuned to novel phonetic characteristics of particular speech sounds (Norris et al. 2003), classes of sounds (Maye et al. 2008, Morley 2008), and even individual words (Dahan & Scarborough 2005). Research on speech production over the past decade has discovered a counterpart to perceptual adaptation. Talkers implicitly imitate the phonetic properties of speech presented in the form of experimental stimuli (Goldinder 1998, Shockley et al. 2004, Nielsen 2007) and of the speech produced by their interlocutors (Pardo 2006, Delvaux & Soquet 2007). The phonetic imitation effect, and its theoretical analysis in terms of a model of phonetic knowledge and learning, is the focus of our paper. We begin by discussing a phonetic imitation experiment that replicates and extends the results of previous work. Participants in the experiment were exposed to speech in which the voice-onset time (VOT) of one word-initial voiceless stop (namely, [p h ]) had been digitally lengthened. When the participants later produced the same words that they had heard as stimuli, and different words beginning with [p h ] that had not heard, their VOTs were longer in comparison to a pre-listening baseline condition. Importantly, the participants also lengthened VOT when producing words that begin with a different voiceless stop, namely [k h ]. This novel result provides evidence against models of speech perception and production that have no representations other than individual sounds or individual words. If such models were correct, participants in this experiment would have had to somehow ‘imitate’ phonetic manipulations that they had not experienced. We analyze phonetic imitation in general, and the extension of imitation from [p h ] to [k h ] in particular, with a model that has multiple levels of linguistic representation and a statistically sound mechanism of adapting to experience. For the purposes of this paper, we take the set of levels to include at least word and feature (or gesture) representations. It is the existence of featural/gestural representations that supports generalization across segments: [p h ] and [k h ] share a feature, such as [– voice] or [+spread glottis] ([+s.g.]), that is phonetically interpreted in terms of VOT (among other phonetic properties; Liberman et al. 1958; Lisker & Abramson 1964); informally, then, we expect a manipulation of the VOT of some [p h ]-initial words to be extended to [k h ]-initial words — that is, to lead to phonetic imitation at the level of the feature — as well as to other [p h ]-initial words. Because the level of phonetic imitation is somewhat greater for words that were actually heard with lengthened VOT (see also Goldinger 1998), we characterize the overall pattern as multi-level phonetic imitation. All else being equal, we expect that all levels of linguistic representation are engaged in phonetic imitation, and therefore that the size of the imitation effect for a particular word will roughly correlate with its similarity, as defined by the model, to words that have been recently heard.