Computing phonological generalization over real speech exemplars

Abstract

We present an Exemplar-Theoretic confidence-sensitive dynamic programming model of speech production, PEBLS (Phonological Exemplar-Based Learning System), and test it with real acoustic speech signals. We focus on the computational problem of how to generate an output that generalizes over a collection of unique, variable-length signals, without resorting to a priori phonological units such as phones or syllables. We show that PEBLS displays pattern-entrenchment behaviour, central to Exemplar Theory's account of phonologization.

Word-types, not word-tokens, facilitate extraction of phonotactic sequences by adults

“…This expectation matches a description of the tokens only condition, however, where no priming was observed. Furthermore, exemplar models do not always store exemplars based on abstract categories, such as phonemes (Goldinger 1998;Kirchner and Moore 2008;but cf. Bailey and Hahn 2001), so it is unclear why word-types would lead to priming in an associative learning model.…”

The Influence of Word Retrieval and Planning on Phonetic Variation: Implications for Exemplar Models

“…Observations such as these have fueled the development of exemplar models of speech production (e.g., Goldinger, 1998; Kirchner, Moore, & Chen, 2010; Pierrehumbert, 2002, 2006; Walsh, Möbius, Wade, & Schütze, 2010; Wedel, 2007; see Pisoni & Levi, 2007, for discussion of similar developments in speech perception and word recognition). In such models, speakers store a large set of rich memory representations, each of which encodes several dimensions of linguistic structure.…”

Modelling the acquisition and evolution of vowel harmony

“…continued necessity of "traditional" phonology There is a common perception among phonologists that instance-based accounts are posed as replacements of traditional phonological analyses, and indeed some exemplar theorists clearly espouse this position(Silverman, 2006;Kirchner and Moore, 2009). Taking seriously the Marrian multi-level view leads us to conclude that this position is likely a mistaken one, and indeed that is what I claim here.…”

“…Coleman (1998,2002) and Phillips (2001) In light of the preceding arguments, I take the type of LIbPhon's instance space to be real-valued and ranging over acoustic features. In particular, I use acoustic formants-bands of increased energy in the spectral representation of human speech (Johnson, 2003)-as the embedding dimension for the simulations discussed here (see Johnson, 1997a;Kirchner and Moore, 2009, for other approaches), as it has been shown that speech tokens synthesised purely from formant information can reliably be recognised by humans (Cooper et al, 1952), and the low dimensionality of this representation makes the simulations more computationally tractable. Vowels are specified by their midpoint formant values and consonants are specified by so-called "locus" values, which can be identified by inspecting the trajectories of consonant-vowel transitions in speech (Lindblom, 1963;Sussman et al, 1998).…”

“…This variability is one of the biggest obstacles in the development of instance-based models of speech production, due to the difficulty of aligning variable-length forms.The algorithms that exist for aligning variable-length real-valued sequences typically involve sophisticated dynamic programming algorithms (e.g. dynamic time warping and hidden Markov modelsRabiner and Juang, 1993), the details of which are not particularly cognitively plausible Kirchner and Moore (2009). use dynamic time warping in their model, PEBLS (2.2.2); an instance-based production-oriented model of spirantisation using real (viz.…”