UC Berkeley

With great effort, adults can try to produce new language sounds with varying degrees of success, yet these same adults automatically and routinely adjust their speech production to accommodate their environments and interlocutors.

This adjustment process allows talkers to be heard above traffic, speak clearly to foreigners, and speak intelligibly with pens in their mouths. All of these actions require monitoring auditory feedback, listening to one's outgoing speech and ensuring that it is error-free. This dissertation investigates the mechanism that allows this automatic maintenance to proceed.

The four experiments described here use a device that systematically alters auditory feedback in real time. Subjects in these experiments hear a slightly distorted version of their own voices, in which one or two of their vowel formants have been shifted. Talkers automatically adjust to these manipulations by opposing the shift in auditory feedback; if they hear their vowels with an artificially high first formant, they start producing vowels with lower first formants. That is, talkers change their speech production to compensate for the shifted auditory feedback.

The first experiment demonstrates that talkers do not change their speech in the way one might naively expect; they do not produce a vowel change exactly inverse to what they hear. The remaining three experiments attempt to characterize what subjects are optimizing when they compensate for shifts in auditory feedback, with particular attention to whether language-specific sound and word patterns constrain automatic speech processing.

In particular, they ask whether low-level compensation for altered auditory feedback is influenced by top-down information from a talker's phonological or lexical inventory, or by acoustic familiarity.

Experiments 3 and 4 find that talkers seek to avoid confusability, compensating less in regions of vowel space with multiple competing vowels, and choose compensation routes that run through acoustically familiar regions of vowel space. Experiment 2 fails to find consistent evidence of influence from the lexical inventory. Together, these experiments demonstrate that even automatic speech processes that operate on low-level auditory information are influenced by high-level knowledge about one's phonological inventory and social context. Results of these experiments are used to expand models of speech motor control and to unify them with psycholinguistic models of speech production.