Neural control of lexical tone production in human laryngeal motor cortex

Vocal pitch is a crucial acoustic cue in spoken language, conveying both lexical and non-lexical information. In non-tonal languages like English, pitch signifies prosody and intonation. However, in tonal languages, spoken by nearly one-third of the world's population, pitch distinguishes words with different meanings. Mandarin Chinese, for example, uses four distinct pitch contours (lexical tones) for a single syllable, each representing a different character and meaning. Precise laryngeal control is critical for producing these pitch variations within the timeframe of a single syllable. The larynx plays a vital role, with three key functions: voicing, pitch rising, and pitch lowering. Voicing, or phonation, results from vocal cord vibration, creating the fundamental frequency (F0). The cricothyroid (CT) and thyroarytenoid (TA) muscles finely control vocal fold tension, influencing pitch. CT stretches the vocal folds, raising pitch, while TA shortens them, lowering pitch. Neuroimaging studies have identified laryngeal motor cortex (LMC) regions in the ventral sensorimotor cortex (vSMC) correlated with laryngeal movements. While the dorsal LMC (dLMC) has been linked to pitch rising in non-tonal languages, the precise neural control mechanisms for generating lexical tones in tonal languages remained unclear. This study aimed to address this gap by investigating the neural mechanisms underlying the precise pitch control of lexical tones in Mandarin speakers.

Prior neuroimaging and neurophysiological studies have identified two LMC regions in the human vSMC associated with laryngeal movements. The bilateral dLMC showed monotonic encoding of pitch rising during a word emphasis task in English. Distinct neural populations in both dLMC and ventral LMC (vLMC) encode voicing. However, how humans precisely control laryngeal muscles to dynamically regulate vocal pitch for lexical tones remained poorly understood. Previous studies, often limited by low temporal resolution, failed to capture the fine-scale neural coding supporting dynamic vocal pitch control. Studies on intonation in non-tonal languages did not fully address the complete range of pitch dynamics, particularly the pitch lowering crucial for tones like the dipping tone in Mandarin. This study aimed to use high-density electrocorticography (ECOG) to overcome temporal resolution limitations and investigate lexical tone production in Mandarin, a typical tonal language, to address this gap.

Eight Mandarin-speaking patients undergoing awake brain surgery for brain tumor resection participated. High-density ECOG grids (128 channels) were temporarily placed on their sensorimotor cortex to record local field potentials during two tone production paradigms and a sentence production task. In the first paradigm, participants named words presented on a screen; the second used auditory cues. The sentence production task involved reading sentences aloud. Neural signals were recorded at 3052 Hz, preprocessed (noise filtering, artifact removal), and the high-gamma activity (70–150 Hz) was extracted. Principal component analysis (PCA) was applied to the acoustic pitch contours of the lexical tones to reduce dimensionality. Electrode localization was performed using intraoperative navigation system data and preoperative MRI. Speech-responsive electrodes were identified by comparing the high-gamma responses around speech onsets with permuted data. Tone-discriminating electrodes were selected based on one-way ANOVA results comparing high-gamma responses across the four tones. A computational model (modified Fujisaki model) was used to analyze pitch contour generation, extracting tone commands (positive and negative) that drove the pitch dynamics. These commands were then used in an encoding model to predict high-gamma activity. Tuning curves were calculated to identify electrodes with positive or negative tuning to tone commands. Finally, direct electrocortical stimulation (DES) was used during awake surgery to causally test the effects of stimulating the identified pitch-encoding regions in the LMC. The study compared evoked pitch changes with control utterances.

The study revealed that local populations in the bilateral LMC encode articulatory kinematic information for generating the pitch dynamics of lexical tones, not tone categories. Two distinct patterns of population activity in the LMC were identified, one commanding pitch rising and the other pitch lowering. These findings were supported by a computational model that accurately predicted high-gamma activity based on tone commands extracted from a modified Fujisaki model. This model showed high fidelity in reconstructing Mandarin pitch contours using extracted tone commands. The study also revealed that electrodes with positive tuning to tone commands were associated with pitch rising, while those with negative tuning were associated with pitch lowering. Direct electrocortical stimulation (DES) of different local populations in the LMC evoked pitch rising and lowering during tone production, respectively. The tone discriminability from neural activity was strongly correlated with pitch height and pitch change representation. Lexical tones could be decoded from distributed neural activity patterns in the SMC using multivariate pattern analysis. The study found that DES of the dLMC could evoke both pitch rising and lowering. Speech arrest was also observed in the bilateral dLMC, particularly near pitch modulation sites, and in the left vLMC.

This study provides strong evidence for a distributed neural representation of pitch control in the bilateral LMC, with distinct populations responsible for pitch rising and lowering. This contrasts with previous findings in non-tonal languages, which predominantly focused on pitch rising. The high temporal resolution of ECOG allowed for detailed characterization of the rapid pitch changes involved in lexical tone production. The combination of correlational encoding models and causal DES provided a refined map of fine-grain pitch control during speech. The findings have significant implications for brain-computer interface (BCI) technology, suggesting that decoding both articulator movements and pitch dynamics from distributed neural activity could enable veridical speech generation for tonal languages like Mandarin.

This study demonstrates that the bilateral LMC contains distinct neural populations controlling pitch rising and lowering during lexical tone production in Mandarin. The findings expand our understanding of LMC function and its neural coding, indicating potential applications in BCI systems for tonal languages. Future research could investigate single-neuron recordings coupled with electromyography and direct vocal fold visualization to refine the understanding of the precise relationship between neural activity and laryngeal muscle movements.

The study's relatively small sample size and the inability to simultaneously record from both left and right LMC in the same patient limited the ability to observe bidirectional tuning across all participants. Future studies with a larger number of participants are needed to better understand the spatial distributions and individual variations in tone production cortex.