Presentation

Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Poster Slams

Symbolic vs. Gradient Phonemes

Poster B26 in Poster Session B and Reception, Thursday, October 6, 6:30 - 8:30 pm EDT, Millennium Hall

Chao Han¹, Arild Hestvik¹, Ryan Rhodes², William Idsardi³; ¹University of Delaware, ²Rutgers University, ³University of Maryland

Are phonemes discrete symbols consisting only of distinctive features (Chomsky & Halle, 1968; Kazanina et al., 2018) or do phonemes contain gradient statistical information about their phonetic properties (Smolensky, Goldrick, & Mathis, 2014)? We addressed this question utilizing the “varying standards” oddball paradigm (Phillips et al., 2000), which assumes that varying the standards within category recruits an abstract phoneme from long-term memory as the memory trace for deviance detection. The resulting MMN therefore reflects the difference between a phoneme and a phonetic oddball. A corollary of this is that not varying the standards (“single-token standards”) should favor a memory trace consisting of the phonetic representation of the standards; the resulting MMN should therefore reflect a phonetics-based discrimination. Both phoneme theories are consistent with a within-category MMN, because listeners can discriminate between phonetic variants and detect phonetic gradience (McMurray et al., 2002; Miglietta et al., 2013). However, the symbolic theory predicts that when standards are varied, the MMN should disappear for a within-category deviant (because the phoneme is a proper subset of the phonetic representation). On the other hand, gradient phoneme theories predict a within-category phonetic MMN even with varying standards, as long as the deviant is an outlier relative to the probability distribution of the phoneme. Experiment 1 compared within-category MMN obtained with varying standards vs. single-token standards (as a between-subjects variable). The varying standards condition (N=33) used 42, 48, 55ms VOT [tæ] stimuli. The single-token condition (N=30) used a 48ms VOT [tæ]. Both conditions used a 119ms VOT [tæ] deviant. MMN was computed as the difference in brain response to the deviant minus the same token in a roving standards control condition. MMN was observed in both conditions/groups (p = .023), and was not modulated by group (p = .478). This rules out the strict interpretation of Phillips et al., (2000) that varying standards recruits a purely symbolic phoneme representation, but is consistent with a theory where the phoneme representation of /t/ contains a Gaussian distribution with a mean VOT of, say, 60ms for [tæ]. In Experiment 2, we asked whether the MMN with varying standards is partially driven by an ad-hoc statistical summary of the standards in addition to experience-based encoding of acoustic statistics. Garrido et al. (2013, 2016) demonstrated with pure tones that MMN to an outlier/deviant is modulated by the statistical distribution of the presented standard stimuli. Adapting the design of Garrido et al. (2013), we presented listeners (N=11) with a 128ms VOT [tæ] deviant embedded in a normal distribution of varying standards with a mean VOT of 64ms and a “wide” standard deviation of 15ms. A second group (N=11) was presented with the same deviant and mean standard but with a “narrow” standard deviation of 5ms. This replicated Experiment 1 with a within-category MMN (p< .001), but no difference between the two standard deviation groups was observed. We discuss alternative interpretations of the lack of an effect of different standard deviations of standards, and the theoretical implications of within-phoneme MMN with varying standards.

Topic Areas: Speech Perception, Phonology and Phonological Working Memory