There is mounting evidence that language users are sensitive to the distributional properties of multi-word sequences. Such findings expand the range of information speakers are sensitive to and call for processing models that can represent larger chains of relations. In the current paper we investigate the effect of multi-word statistics on phonetic duration using a combination of experimental and corpus-based research. We ask (a) if phonetic duration is affected by multi-word frequency in both elicited and spontaneous speech, and (b) if syntactic constituency modulates the effect. We show that phonetic durations are reduced in higher frequency sequences, regardless of constituency: duration is shorter for more frequent sequences within and across syntactic boundaries. The effects are not reducible to the frequency of the individual words or substrings. These findings open up a novel set of questions about the interaction between surface distributions and higher order properties, and the resulting need (or lack thereof) to incorporate higher order properties into processing models.

The duration and occasional deletion rate of consonants differ from one language to another. What causes a language to preserve and lengthen some consonants but shorten and delete others? I show that the typology of consonant duration and occasional deletion in American English is affected by consonants' informativity -their average local predictability. Informativity can explain why usually-predictable segments such as American English /t/ are likely to be reduced even when they are locally unpredictable, but usuallypredictable segments are preserved even when they are redundant. I use four corpus studies to demonstrate that higher informativity leads to longer duration and reduced likelihood to delete even when other important factors such as the phonetic features, frequency, and local predictability of consonants are controlled for. The role of informativity in the duration and deletion rates of consonants can bridge the gap between phonetic performance and the actuation of phonological processes.

Speakers have been shown to alter their speech to resemble that of their conversational partner. Do speakers converge with their interlocutor's baseline, or does convergence stem from conversational properties that similarly affect both parties? Using the Switchboard corpus, this paper shows evidence for speakers' convergence in speech rate to the other party's baseline, not only to conversation-specific properties. Study 1 shows that the method for calculating speech rate used in this paper is powerful enough to replicate established findings. Study 2 demonstrates that speakers are mostly affected by their own behavior in other contexts, but that they also converge to their interlocutor's baseline, established using the interlocutor's behavior in other contexts. Study 2 also shows that speakers change their speech rate in response to the interlocutor's characteristics: speakers speak more slowly with older speakers regardless of the interlocutor's speech rate, and male speakers speak faster with other male speakers.