Neural speech decoding with magnetoencephalography

Severe brain damage or amyotrophic lateral sclerosis (ALS) may lead the patients to a locked-in state where the patients are motorically paralyzed otherwise being cognitively normal. The brain might be the only source of communication for these patients. Current commercially available brain-computer interface (BCI) spellers can help these patients communicate to a level but at a very slow communication rate (less than 10 words/min). Neural speech decoding paradigm attempts to decode speech information directly from the brain providing promise towards a faster communication assistance, thereby, improving the quality of life for these patients. Magnetoencephalography (MEG) is a non-invasive neuroimaging modality that has an excellent spatio-temporal resolution, suitable to study neural mechanism of speech. This dissertation, for the first time, investigates the possibility of neural speech decoding using MEG in the following aspects: imagined and overt speech decoding for healthy and ALS individuals, subject generalization in decoding, articulation and acoustic synthesis, and MEG sensor selection for optimal decoding. The possibility of decoding speech from MEG was evident from the findings showcased in this dissertation providing a solid foundation towards future wearable MEG based speech-BCI applications.