Single-neuronal elements of speech production in humans

Humans possess the remarkable ability to produce a vast array of sounds to convey meaning through speech. Fluent speech requires a rapid, structured sequence of processes to plan the arrangement and structure of phonemes within words. These processes are thought to involve prefrontal regions, part of a broader language network implicated in word planning, sentence construction, and motor production. While previous research using cortical surface recordings has indicated regional organization of phonetic features, the fundamental cellular mechanisms underlying word planning and production remain unclear. Previous studies in animals and limited human research have shed light on the relationship between motor areas and vocalization, but they haven't fully elucidated the neuronal processes involved in constructing individual words during natural speech. Linguistic theories suggest tightly coupled sublexical processes coordinating articulators during word planning, but how specific phonetic sequences, their syllabification, and inflection are encoded by individual neurons is largely unknown. Despite significant regional overlap in areas involved in articulation planning and production, the unique cellular representation of these processes and their cortical organization remain poorly understood. Single-neuronal recordings offer a powerful tool to investigate the fundamental functional building blocks of human word planning and production, providing insights at previously inaccessible spatiotemporal scales. This study leverages recently developed ultrahigh-density microelectrode arrays, speech tracking techniques, and advanced modeling approaches to address these critical questions.

Existing literature points to the involvement of prefrontal regions in speech production, particularly in word planning and sentence construction. Studies using fMRI and other neuroimaging techniques have identified brain areas associated with various aspects of speech production, but the cellular level mechanisms have remained elusive. Animal models have provided valuable insights into vocalization control, but these don't directly translate to the complexities of human speech. There is a theoretical basis for the tightly coupled nature of sublexical processes during speech planning, but empirical evidence at a single-neuron level has been lacking. While the overall brain regions involved in planning and production show overlap, there's limited understanding of the unique cellular representations and cortical organization at a finer resolution. This study fills this gap by employing a novel combination of advanced recording techniques and analytical methods.

This study utilized acute, ultrahigh-density Neuropixels recordings from the language-dominant (left) prefrontal cortex of participants undergoing planned intraoperative neurophysiology. Recordings were obtained from the posterior middle frontal gyrus, a region known for its role in word planning, sentence construction, and connections to motor areas involved in articulation. Awake participants performed a naturalistic speech production task, articulating diverse words in a reproducible manner. The task required participants to generate words varying in phonetic, syllabic, and morphosyntactic content, ensuring natural speech production independent of explicit phonetic cues. Controls were included to assess for word-related responses, sensory-perceptual effects, and phonetic-acoustic properties. Ultrahigh-density Neuropixels arrays provided high-throughput recordings of single cortical units, with custom software used for registration and motion correction of action potential (AP) activity. Well-isolated single units with stable waveform morphologies were selected for analysis. A total of 272 putative neurons were recorded across five participants. To examine the relationship between neuronal activity and speech production, a feature space was constructed based on the constituent phonemes of each word. Generalized linear models (GLMs) quantified the degree to which neuronal activity variations during planning could be explained by individual phonemes. Hamming distance calculations assessed the relationship between neuronal activity and specific phoneme combinations. Multilabel decoders were used to classify upcoming phonemes and evaluate decoding accuracy. A perception control task was conducted to compare neuronal activities during speech production and perception. The analysis further investigated neuronal responses to syllables and morphemes, constructing an additional vector space based on syllabic structure and order. Finally, a dynamical systems approach analyzed the transition from word planning to production, identifying functional subspaces and tracking their evolution over time.

The study revealed that a significant proportion (46.7%) of recorded neurons showed activity patterns informative of the phonetic content of words before utterance. A subset of these neurons (20.6%) exhibited selective tuning to specific planned phonemes, encoding information about both place and manner of articulation, and reflecting spectral properties of articulated phonemes. These neurons encoded not only individual phonemes but also their specific combinations within words. Furthermore, the phonetic composition of upcoming words could be reliably decoded from neuronal activity with significant accuracy (ROC-AUC = 0.75). Neurons showing phonetic selectivity during speech planning were largely distinct from those exhibiting selectivity during perception, highlighting the unique neural processing involved in each. Many neurons (25%) reflected the presence of specific planned syllables, accurately predicting syllabic content. These neurons demonstrated selective tuning to specific syllables, even when controlling for constituent phonemes and their order. A smaller proportion of neurons (11.4%) showed selectivity for morphemes, with neuronal activity predicting morpheme inclusion. Neurons encoding sublexical components were broadly distributed across the cortex and cortical column depth, with a slightly higher preponderance at lower depths. The proportion of selective neurons increased posteriorly along the rostral-caudal axis. A notable finding was the observed temporal dynamic: decoding performance peaked for morphemes, followed by phonemes, and then syllables, suggesting a temporally ordered morphological-phonetic-syllabic dynamic. This temporal order was consistent with previous neurolinguistic models, indicating that morphology is likely retrieved before phonology during word production. The study also demonstrated a distinct transition of neuronal activity from articulation planning to production, with neurons encoding similar information during both phases. Dynamical systems analysis revealed that the neural population occupied largely separate subspaces during planning and production, suggesting a mechanism for the rapid separation of neural processes involved in word construction and articulation.

The findings provide compelling evidence for a highly structured organization of phonetic representations in the human prefrontal cortex at a cellular level. The observed encoding of phonemes, their combinations, syllables, and morphemes, alongside their temporal dynamics, supports existing linguistic models proposing tightly coupled sublexical processes during speech planning. The distinction between neural populations involved in speech production versus perception emphasizes the specialized nature of these processes. The observed spatial distribution suggests a functional organization along the cortical column, and a potential for redundancy in phonetic information representation within local cortical populations. The temporal succession of neuronal encoding, revealing a morphological-phonetic-syllabic dynamic, provides insights into the precise sequence of cognitive processes involved in word production. The observed distinct subspaces during planning and production phases provide strong support for a dynamical systems model of speech articulation, suggesting a mechanism that allows for the efficient separation and coordination of neural processes crucial for fluent speech.

This study provides novel insights into the cellular mechanisms underlying human speech production, revealing a remarkably detailed and structured organization of phonetic representations in the prefrontal cortex. The findings highlight the involvement of distinct neuronal populations encoding various sublexical features with a specific temporal order. These results offer a foundation for future research exploring the neural basis of language, potentially informing the development of advanced speech prostheses and brain-machine interfaces. Further research should investigate the involvement of other brain areas and the influence of higher-level linguistic processes such as semantics and prosody.

The study's findings are based on recordings from a relatively small sample of participants and a limited cortical region. The generalizability of the findings across different language speakers and other brain regions requires further investigation. The study focused primarily on phonological aspects of speech production; future research should incorporate semantic and syntactic factors for a more comprehensive understanding. The use of acute intraoperative recordings limits the duration of observation and the range of possible experimental paradigms.