Language, Cognition and Neuroscience ISSN: 2327-3798 (Print) 2327-3801 (Online) Journal homepage: https://www.tandfonline.com/loi/plcp21 Articulating: the neural mechanisms of speech production Elaine Kearney & Frank H. Guenther To cite this article: Elaine Kearney & Frank H. Guenther (2019): Articulating: the neural mechanisms of speech production, Language, Cognition and Neuroscience, DOI: 10.1080/23273798.2019.1589541 To link to this article: https://doi.org/10.1080/23273798.2019.1589541 Published online: 05 Mar 2019. Submit your article to this journal View Crossmark data Full Terms & Conditions of access and use can be found at https://www.tandfonline.com/action/journalInformation?journalCode=plcp21 LANGUAGE, COGNITION AND NEUROSCIENCE https://doi.org/10.1080/23273798.2019.1589541 REGULAR ARTICLE Articulating: the neural mechanisms of speech production Elaine Kearneya and Frank H. Guenthera,b,c,d aDepartment of Speech, Language, and Hearing Sciences, Boston University, Boston, MA, USA; bDepartment of Biomedical Engineering, Boston University, Boston, MA, USA; cThe Picower Institute for Learning and Memory, Massachusetts Institute of Technology, Cambridge, MA, USA; dAthinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA, USA ABSTRACT ARTICLE HISTORY Speech production is a highly complex sensorimotor task involving tightly coordinated processing Received 12 October 2018 across large expanses of the cerebral cortex. Historically, the study of the neural underpinnings of Accepted 21 February 2019 speech suffered from the lack of an animal model. The development of non-invasive structural and KEYWORDS functional neuroimaging techniques in the late twentieth century has dramatically improved our fl Articulation; motor control; understanding of the speech network. Techniques for measuring regional cerebral blood ow neuroimaging; fMRI; PET; have illuminated the neural regions involved in various aspects of speech, including feedforward meta-analysis and feedback control mechanisms. In parallel, we have designed, experimentally tested, and refined a neural network model detailing the neural computations performed by specific neuroanatomical regions during speech. Computer simulations of the model account for a wide range of experimental findings, including data on articulatory kinematics and brain activity during normal and perturbed speech. Furthermore, the model is being used to investigate a wide range of communication disorders. 1. Introduction system from early work with non-human primates and Speech production is a highly complex motor act invol- more recently with neural modelling and experimental ving respiratory, laryngeal, and supraglottal vocal tract testing. articulators working together in a highly coordinated The present article takes a historical perspective in fashion. Nearly every speech gesture involves several describing the neural mechanisms of speech motor articulators – even an isolated vowel such as “ee” control. We begin in the first section with a review of involves coordination of the jaw, tongue, lips, larynx, models and theories of speech production, outlining and respiratory system. Underlying this complex motor the state of the field in 1989 and introducing the DIVA act is the speech motor control system that readily inte- model – a computational neural network that describes grates auditory, somatosensory, and motor information the sensorimotor interactions involved in articulator represented in the temporal, parietal, and frontal control during speech production (Guenther, 1995; cortex, respectively, along with associated sub-cortical Guenther, Ghosh, & Tourville, 2006). Taking a similar structures, to produce fluent and intelligible speech – approach in the next section, we review the key empiri- whether the speech task is producing a simple nonsense cal findings regarding the neural bases of speech pro- syllable or a single real word (Ghosh, Tourville, & duction prior to 1989 and highlight the primary Guenther, 2008; Petersen, Fox, Posner, Mintun, & developments in cognitive neuroimaging that followed Raichle, 1988; Sörös et al., 2006; Turkeltaub, Eden, and transformed our ability to conduct non-invasive Jones, & Zeffiro, 2002). speech research in humans. The neural correlates of In Speaking, Levelt (1989) laid out a broad theoretical speech production are discussed in the context of the framework of language production from the conceptual- DIVA model; as a neural network, the model’s com- isation of an idea to the articulation of speech sounds. In ponents correspond to neural populations and are comparison to linguistic processes, speech motor control given specific anatomical regions that can then be mechanisms differ in a number of ways. They are closer tested against neuroimaging data. Data from exper- to the neural periphery, more similar to neural substrates iments that investigated the neural mechanisms of audi- in non-human primates, and better understood in terms tory feedback control are presented to illustrate how the of neural substrates and computations. These character- model quantitatively fits to both behavioural and neural istics have shaped the study of the speech motor control data. In the final section, we demonstrate the utility of CONTACT Frank H. Guenther [email protected] © 2019 Informa UK Limited, trading as Taylor & Francis Group 2 E. KEARNEY AND F. H. GUENTHER neurocomputational models in furthering the scientific To overcome this problem, several models postulate understanding of motor speech disorders and informing that auditory targets may be equated to targets in a the development of novel, targeted treatments for those somatosensory reference frame that is more closely who struggle to translate their message “from intention related to the articulators than an acoustic reference to articulation”. frame (e.g. Lindblom, Lubker, & Gay, 1979; Perkell, 1981). For example, Lindblom et al. (1979) proposed that the area function of the vocal tract (i.e. the 3D 2. Models and theories of speech production shape of the vocal tract “tube”), which largely determines In summarising his review of the models and theories of its acoustic properties, acts as a proxy for the auditory speech production, Levelt (1989, p. 452) notes that target that can be sensed through somatic sensation. “There is no lack of theories, but there is a great need Furthermore, they posit that the brain utilises an internal of convergence.” This section first briefly reviews a model that can estimate the area function based on number of the theoretical proposals that led to this con- somatosensory feedback of articulator positions and clusion, culminating with the influential task dynamic generate corrective movements if the estimated area model of speech production, which appeared in print function mismatches the target area function. the same year as Speaking. We then introduce the DIVA Published in the same year as Levelt’s landmark book, model of speech production, which incorporates many the task dynamic model (Saltzman & Munhall, 1989) pro- prior proposals in providing a unified account of the vided a fleshed-out treatment of vocal tract shape neural mechanisms responsible for speech motor targets. According to this model, the primary targets of control. speech are the locations and degrees of key constrictions of the vocal tract (which dominate the acoustic signal compared to less-constricted parts of the vocal tract), 2.1. State of the field prior to 1989 specified within a time-varying gestural score. The One of the simplest accounts for speech motor control is model was mathematically specified and simulated on the idea that each phoneme is associated with an articu- a computer to verify its ability to achieve constriction latory target (e.g. MacNeilage, 1970) or a muscle length targets using different combinations of articulators in target (e.g. Fel’dman, 1966a, 1966b) such that production different speaking conditions. of the phoneme can be carried out simply by moving the The task dynamic model constitutes an important articulators to that muscle/articulatory configuration. By milestone in speech modelling and continues to be 1989, substantial evidence against such a simple articula- highly influential today. However, it does not account tory target view was already available, including studies for several key aspects of speech: for example, the indicating that, unlike some “higher-level” articulatory model is not neurally specified, it does not account for targets such as lip aperture, individual articulator pos- development of speaking skills (all parameters are pro- itions often vary widely for the same phoneme depend- vided by the modeller rather than learned), and it does ing on things like phonetic context (Daniloff & Moll, 1968; not account for auditory feedback control mechanisms Kent, 1977; Recasens, 1989), external loads applied to the such as those responsible for compensatory responses jaw or lip (Abbs & Gracco, 1984; Folkins & Abbs, 1975; to purely auditory feedback manipulations. The DIVA Gracco & Abbs, 1985), and simple trial-to-trial variability model introduced in the following subsection addresses (Abbs, 1986). these issues by integrating past proposals such as audi- The lack of invariance in the articulator positions used tory targets, somatosensory targets, and internal to produce phonemes prompted researchers to search models into a relatively straightforward, unified for a different kind of phonemic “target” that could account of both behavioural and neural