Language represents one of the most complex and, at the same time, fundamental cognitive functions in human life. However, certain neurological disorders can cause paralysis conditions that severely compromise communication with the external world, both verbal and gestural. To preserve quality of life under these conditions, various forms of Augmentative and Alternative Communication (AAC) have been developed. Nevertheless, currently available technologies present significant limitations, as they rely on slow and unintuitive paradigms, often unsuitable for the most severe cases of paralysis. In recent years, there has been growing interest in the development of speech brain-computer interfaces (speech BCIs), neural interfaces capable of decoding language directly from brain activity. The paradigm employed in such applications is speech imagery, i.e., the imagination of speech without overt sound production or articulatory movements. This approach offers greater intuitiveness and speed of use, making it a promising alternative to classical approaches based on event-related potentials or motor imagery. The present work focuses on the characterization of the neurophysiological correlates of speech imagery through the analysis of electroencephalographic (EEG) signals, with the ultimate goal of contributing to the development of a non-invasive speech BCI. After a brief overview of the main neurophysiological models of speech production and perception, as well as the current state of the art in speech BCIs, an experimental study is presented in which healthy participants performed overt and silent speech tasks involving sentences and isolated syllables. EEG signals were preprocessed using standard cleaning procedures, including spatial and frequency filtering, artifact removal through ASR and ICA, interpolation of noisy channels, and rejection of contaminated trials. Time-domain analysis of event-related potentials (ERPs) was then conducted. To reduce the risk of waveform cancelation due to phonetic variability, tasks were classified according to syllabic structure and articulatory features. The analyses revealed the presence of evoked potentials related to the experimental protocol. In contrast, in brain regions typically associated with linguistic functions, the observed activity does not appear to be interpretable as ERPs linked to syllabification during speech imagery. It is possible that the absence of strong evidence is due to the protocol employed and to the intrinsic complexity of the phenomenon, suggesting the need for future studies with more targeted experimental approaches.
Il linguaggio rappresenta una delle funzioni cognitive più complesse e al tempo stesso fondamentali nella vita umana. Alcune patologie neurologiche possono però portare a condizioni di paralisi che compromettono la comunicazione con l’esterno, sia verbale che gestuale. Al fine di garantire una buona qualità di vita alle persone colpite da questa condizione sono state sviluppate forme di comunicazione alternativa(AAC). Le tecnologie ad oggi disponibili presentano tuttavia numerose limitazioni, in quanto si basano su paradigmi lenti, poco intuitivi e spesso non adatti ai casi più gravi di paralisi. Negli ultimi anni è emerso un crescente interesse per lo sviluppo di ”speech brain computer interfaces” (speech BCI), interfacce neurali in grado di decodificare il linguaggio sfruttando i segnali neurali dell’utente. Il paradigma impiegato da questo tipo di applicazioni è lo ”speech imagery”, ovvero l’immaginazione del parlato senza produzione di suono o movimento. Tale approccio garantisce maggiore intuitività e velocità di utilizzo, rendendolo una potenziale alternativa agli approcci classici basati su potenziali evocati o ”motor imagery”. Il presente elaborato si concentra sulla caratterizzazione dei correlati neurofisiologici dello ”speech imagery” attraverso l’analisi di segnali elettroencefalografici, con l’obiettivo finale di contribuire allo sviluppo di una speech BCI non invasiva. Dopo una breve rassegna dei principali modelli neurofisiologici della produzione e comprensione del parlato, e dello stato dell’arte delle speech BCI, viene presentato uno studio sperimentale in cui partecipanti sani hanno eseguito compiti di parlato in condizione ”overt” e ”silent” di frasi e sillabe isolate. I segnali EEG sono stati processati con procedure standard di pulizia, inclusi filtraggio spaziale e in frequenza, rimozione del rumore tramite ASR e ICA, interpolazione dei canali rumorosi ed eliminazione dei trial contaminati da artefatti residui. Successivamente è stata condotta un’analisi nel dominio temporale dei potenziali evento-correlati (ERP). Per ridurre il rischio di cancellazione dei segnali a causa della variabilità fonetica, i task sono stati classificati sulla base della struttura sillabica e delle caratteristiche articolatorie (consonanti occlusive, fricative, nasali; vocali alte, medie, basse). Le analisi hanno evidenziato la presenza di potenziali evocati legati al protocollo sperimentale. Nelle regioni tipicamente associate alle funzioni linguistiche, invece, l’attività osservati non appare interpretabile come ERP riconducibili alla sillabazione in speech imagery. È possibile che l’assenza di forti evidenze sia dovuta al protocollo adottato e alla complessità intrinseca del fenomeno, indicando la necessità di studi futuri con approcci sperimentali più mirati.
Analysis of neural correlates of speech imagery for non-invasive brain-computer interfaces
GNOCATO, MARGHERITA
2024/2025
Abstract
Language represents one of the most complex and, at the same time, fundamental cognitive functions in human life. However, certain neurological disorders can cause paralysis conditions that severely compromise communication with the external world, both verbal and gestural. To preserve quality of life under these conditions, various forms of Augmentative and Alternative Communication (AAC) have been developed. Nevertheless, currently available technologies present significant limitations, as they rely on slow and unintuitive paradigms, often unsuitable for the most severe cases of paralysis. In recent years, there has been growing interest in the development of speech brain-computer interfaces (speech BCIs), neural interfaces capable of decoding language directly from brain activity. The paradigm employed in such applications is speech imagery, i.e., the imagination of speech without overt sound production or articulatory movements. This approach offers greater intuitiveness and speed of use, making it a promising alternative to classical approaches based on event-related potentials or motor imagery. The present work focuses on the characterization of the neurophysiological correlates of speech imagery through the analysis of electroencephalographic (EEG) signals, with the ultimate goal of contributing to the development of a non-invasive speech BCI. After a brief overview of the main neurophysiological models of speech production and perception, as well as the current state of the art in speech BCIs, an experimental study is presented in which healthy participants performed overt and silent speech tasks involving sentences and isolated syllables. EEG signals were preprocessed using standard cleaning procedures, including spatial and frequency filtering, artifact removal through ASR and ICA, interpolation of noisy channels, and rejection of contaminated trials. Time-domain analysis of event-related potentials (ERPs) was then conducted. To reduce the risk of waveform cancelation due to phonetic variability, tasks were classified according to syllabic structure and articulatory features. The analyses revealed the presence of evoked potentials related to the experimental protocol. In contrast, in brain regions typically associated with linguistic functions, the observed activity does not appear to be interpretable as ERPs linked to syllabification during speech imagery. It is possible that the absence of strong evidence is due to the protocol employed and to the intrinsic complexity of the phenomenon, suggesting the need for future studies with more targeted experimental approaches.| File | Dimensione | Formato | |
|---|---|---|---|
|
Gnocato_Margherita.pdf
accesso aperto
Dimensione
2.66 MB
Formato
Adobe PDF
|
2.66 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/99045