Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Continuous-Time Recurrent Neural Network Models of Speech Motor Cortex Suggest Stress-Specific Word Motor Programs
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Latane Bullock1,2, Ram Zaveri3, Julia Chauvet4, Mélen Guillaume5,6, Frank Guenther7,8; 1Program in Speech and Hearing Bioscience and Technology, Division of Medical Sciences, Harvard Medical School, Boston, MA, 02115, USA, 2Department of Neurosurgery, Massachusetts General Hospital, Boston, MA, 02114, USA, 3Department of Biomedical Engineering, Boston University, Boston, MA, United States, 4Max Planck Society for the Advancement of the Sciences - Max Planck Institute for Psycholinguistics, 5Univ. Lille, CNRS, UMR 9193 - SCALab - Sciences Cognitives et Sciences Affectives, F-59000 Lille, France, 6Univ. Grenoble Alpes, Univ. Savoie Mont Blanc, CNRS, LPNC, 38000 Grenoble, France, 7Department of Speech, Language, and Hearing Sciences, Boston University, Boston, MA, United States, 8Picower Institute for Learning and Memory, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
Computational models have advanced our understanding of the neural basis of speech production, but their temporal resolution has been limited by their reliance on fMRI data. Many existing models predict relative activity levels across brain regions while remaining agnostic to within-region neural dynamics. Converging evidence suggests that modeling speech motor cortex as a high-dimensional dynamical system, with key computations unfolding in lower-dimensional latent subspaces, may provide additional explanatory power (Guenther, Chang, Cannon 2026 PsyArXiv). Continuous-time artificial recurrent neural networks (CT-RNNs) have provided influential accounts of motor cortical dynamics during arm reaching in non-human primates (Churchland…Shenoy 2012 Nature et seq). Here, we translate this dynamical systems framework to speech production. We trained CT-RNNs with 100 hidden units and linear readouts to transform time-varying speech inputs (representing phoneme sequences and words) into time-varying articulatory and laryngeal outputs derived from electromagnetic articulography. The networks learned articulatory trajectories with high accuracy, sufficient to generate intelligible speech waveforms using a speech articulatory-to-acoustic inversion model. We then used this framework to test an open question in speech neuroscience: are well-learned word-level motor programs stress-specific or stress-invariant? That is, does the speech motor system maintain distinct representations for forms such as ‘obJECT’ and ‘OBject’, or a single representation for ‘object’ modulated by an external stress signal? We used a preexisting dataset of Dutch speakers producing lexical stress minimal pairs to probe this question (Severijnen et al. 2024). CT-RNNs trained with stress-specific word inputs produced significantly more accurate articulatory trajectories than networks given stress-invariant word inputs modulated by an external stress signal (Wilcoxon, p < 0.001), despite having identical numbers of free parameters. Furthermore, this advantage was not explained by differences in rate of learning (p = 0.89). These findings are evidence for a computational advantage of stress-specific word motor programs, and suggest that speech motor cortex may encode a stress-specific repertoire of well-learned word forms. The neurocomputational motifs underlying chunked, hierarchical, goal-driven movements like speech are only beginning to be unraveled. The proliferation of single-neuron recordings during speech warrants a constrained model-data loop to distill and probe high-dimensional recordings. Our work provides a foundation for future comparisons between modeled and recorded neural dynamics during speech.
Topic Areas: Computational Approaches, Speech Motor Control