Languages vary in their phonotactics: how sounds are allowed to be sequenced into words. For example, although English, Spanish, and Farsi all have the sounds /p l s/, English allows the complex onsets /sp/ and /pl/, as in ‘speech’ and ‘play’, Spanish only /pl/, as in ‘playa’, while Farsi allows no complex onsets at all. Learners of a second language (L2) often struggle with producing complex onsets that are banned in their first language (L1) and will commonly repair them using epenthesis: the insertion of a vowel to break up an illegal sequence. There is a robust asymmetry in repair strategies for complex onsets across languages: sC onsets are more typically repaired by epenthesizing before the cluster (prothesis; as in ‘stop’ → ‘estop’) while other onsets are repaired by epenthesizing into the cluster (anaptyxis; as in ‘please’ → ‘pelease’). One analysis of this asymmetry has been in terms of sonority, the overall resonance of a sound: this proposal suggests that there is a pressure to avoid sonority rises across syllable boundaries (Vennemann 1988, Gouskova 2001, 2004). A second analysis suggests that speakers choose epenthesis locations that maximize perceptual similarity to the unepenthesized form (Fleischhacker 2001, 2005, Flemming 2008).

In this talk, I will compare the predictions of these two theories against new empirical data from L1 Farsi/L2 English speakers. In the first part of the talk, I will present the results of a corpus study and an experimental study, which consist of recordings of speakers of varying English proficiency reading English passages. The results of these studies confirm previous observations that (a) sC onsets almost categorically undergo prothesis, while other onsets undergo anaptyxis; (b) overall rates of cluster repair decrease as English ability increases; and (c) sC clusters are repaired more frequently and acquired more slowly than other clusters. In the second part of the talk, I will present a maximum entropy optimality theory model of the experimental data. First, I will show that constraint scaling can be used to model the effects of L2 proficiency on epenthesis paterns, providing a starting point for more nuanced modeling of
L2 acquisition. Second, I will demonstrate that when the models are extended to encode the increased difficulty of sC clusters, the perceptual account of epenthesis asymmetry beter predicts the data than the syllable contact account. Although the perceptual model makes errors in the overall rates of epenthesis, it consistently predicts the correct type. The syllable contact model, on the other hand, fails to predict both the rate and type of epenthesis of many clusters. I will close by arguing that models that treat sonority as a scalar value, such as the syllable contact account or the more recent NAP model (Albert & Nicenboim 2022), are ultimately insufficient to deal with the wide range of sonority-related phenomena observed in speech. I suggest instead that further study of the articulatory and perceptual reflexes of sonority will be important to ensure our models are both empirically and explanatorily adequate. Specifically, I will suggest that the concept of ‘good’ vs. ‘bad’ variability (e.g. Latash 2012) from the motor control literature may provide a principled way to relate the perceptual and articulatory properties of sC clusters. – Connor Mayer