Robust Sensorimotor Learning during Variable Sentence-Level Speech

Abstract

Sensorimotor learning has been studied by altering the sound of the voice in real time as speech is produced. In response to voice alterations, learned changes in production reduce the perceived auditory error and persist for some time after the alteration is removed [1-5]. The results of such experiments have led to the development of prominent models of speech production. This work proposes that the control of speech relies on forward models to predict sensory outcomes of movements, and errors in these predictions drive sensorimotor learning [5-7]. However, sensorimotor learning in speech has only been observed following intensive training on ahandful of discrete words or perceptually similar sentences. Stereotyped production does not capture the complex sensorimotor demands of fluid, real-world speech [8-11]. It remains unknown whether talkers predict the sensory consequences of variable sentence production to allow rapid and precise updating of speech motor plans when sensory prediction errors are encountered. Here, we used real-time alterations of speech feedback to test for sensorimotor learning during the production of 50 sentences that varied markedly in length, vocabulary, and grammar. Following baseline production, all vowels were simultaneously altered and played back through headphones in near real time. Robust feedforward changes in sentence production were observed that, on average, precisely countered the direction of the alteration. These changes occurred in every participant and transferred to the production of single words with varying vowel sounds. The results show that to maintain accurate sentence production, the brain actively predicts the auditory consequences of variable sentence-level speech.