Allophones, not phonemes in spoken-word recognition

Abstract

•Pre-lexical representations in speech perception were probed with selective adaptation • Allophonic and phoneme overlap between adaptors and test stimuli were varied • Only allophonic overlap led to selective adaptation • Results argue for the use of allophones and not phonemes in speech perceptionWhat are the phonological representations that listeners use to map information about the segmental content of speech onto the mental lexicon during spoken-word recognition? Recent evidence from perceptual-learning paradigms seems to support (context-dependent) allophones as the basic representational units in spoken-word recognition. But recent evidence from a selective-adaptation paradigm seems to suggest that context-independent phonemes also play a role. We present three experiments using selective adaptation that constitute strong tests of these representational hypotheses.In Experiment 1, we tested generalization of selective adaptation using different allophones of Dutch /r/ and /l/ -a case where generalization has not been found with perceptual learning. In Experiments 2 and 3, we tested generalization of selective adaptation using German back fricatives in which allophonic and phonemic identity were varied orthogonally. In all three experiments, selective adaptation was observed only if adaptors and test stimuli shared allophones. Phonemic identity, in contrast, was neither necessary nor sufficient for generalization of selective adaptation to occur. These findings and other recent data using the perceptuallearning paradigm suggest that pre-lexical processing during spoken-word recognition is based on allophones, and not on context-independent phonemes.

Feature Acquisition in Second Language Phonetic Development: Evidence From Phonetic Training

“…As an example, some researchers have interpreted findings from selective adaptation as indicating processing based on acoustic similarity rather than on abstract phonological units (for selective adaptation without phonemic overlap, see Goldinger, Luce, & Pisoni, 1989; for selective adaptation in nonspeech sounds, see Remez, 1979). Another subphonemic approach can be found in word recognition research, with some suggesting that context-dependent sublexical units (Reinisch, Wozny, Mitterer, & Holt, 2014) such as allophones (Mitterer, Reinisch, & McQueen, 2018) form the basis for spoken word recognition. Eschewing the traditional fundamental unit debate, Goldinger and Azuma (2003) posit that units may be reconceptualized as self-organizing dynamic states, although again this does not preclude links between different phonemelike units.…”

Phoneme‐Order Encoding During Spoken Word Recognition: A Priming Investigation

“…In this study, we provide a further test of a role for position‐independent phonemes in spoken word recognition by using a priming paradigm. This is an important endeavor because, to the best of our knowledge, they are only two studies to date that have examined this important question (Gregg, Inhoff, & Connine, in press; Toscano et al, ), and also because Toscano et al's findings were recently questioned regarding their statistical reliability (Mitterer, Reinisch, & McQueen, ). It is also important because, to our knowledge, there is only one model of spoken word recognition, the TISK model (Hannagan et al, ), that can account for Toscano et al's findings.…”

Machine Learning–based Analysis of English Lateral Allophones

“…The ANN algorithm has an excellent generalization capability to learn from the set of data and has been used for decades in automatic speech recognition. Moreover, recently ANNs have been employed in automatic phoneme recognition, hence the research very closely related to allophone extraction and classification (Mitterer et al, 2018;Kozierski et al, 2016).…”

“…Such a comparison cannot, however, be performed in a straightforward way, as other works in this area concentrate around speech recognition and not on elements of speech, such as allophones (Mitterer et al, 2018;Ali et al, 1999;Kozierski et al, 2016;Baghdasaryan and Beex, 2011). Kozierski and his collaborators came to the conclusion that an approach based on allophones cannot directly be used in automatic speech recognition without further research and modification of the employed methods (Kozierski et al, 2016).…”