Perceptual similarity in input–output mappings: A computational/experimental study of non-native speech production

This paper takes a computational/experimental approach to investigating faithfulness in input–output phonological mappings. We seek to explain the results of a speech production experiment recently reported in Davidson (2010). In that experiment, native English speakers were asked to produce phonotactically unattested consonant clusters. We argue that modifications of the target consonant clusters are best understood by considering both unfaithful phonological mappings and imprecision in the speech production mechanism. To account for the pattern of unfaithful input–output mappings, we consider an extension of the P-map hypothesis (Steriade, 2008) to the production of phonotactically unattested target sequences. Predictions of the P-map for this data were established by a perception experiment in which participants were asked to discriminate between unattested consonant clusters, CC, and attested modifications, including epenthesis, prothesis, C1 change and C1 deletion. To evaluate the effects of motor noise on consonant cluster production, we constructed a computational model that allowed us to simulate consonant cluster productions under different levels of noise. Simulations reveal that, first, a large proportion of cases involving vocoid insertion, CəC, are better accounted for by noisy implementation of the target timing than by phonological epenthesis and, second, that differences in insertion patterns between stop-initial clusters and fricative-initial clusters are due to the internal temporal properties of stops and fricatives. Factoring these cases into the analysis isolates a pattern of unfaithful input–output mappings used to evaluate the P-map hypothesis. On the basis of considerable mismatches between patterns of perceptual similarity and unfaithful input–output mappings, we argue that the P-map theory of faithfulness is too restrictive to extend to the production of non-native speech.