Decoding Syllables from Human fMRI Activity

Abstract

Language plays essential roles in human cognition and social communication, and therefore technology of reading out speech using non-invasively measured brain activity will have both scientific and clinical merits. Here, we examined whether it is possible to decode each syllable from human fMRI activity. Four healthy subjects participated in the experiments. In a decoding session, the subjects repeatedly uttered a syllable presented on a screen at 3Hz for a 12-s block. Nine different syllables are presented in a single experimental run which was repeated 8 times. We also specified the voxels which showed articulation-related activities by utterance of all the syllables in Japanese phonology in a conventional task-rest sequence. Then, we used either all of these voxels or a part of these voxels that exist in anatomically specified ROIs (M1, cerebellum) during decoding sessions as data samples for training and testing a decoder (linear support vector machine) that classifies brain activity patterns for different syllables. To evaluate decoding performance, we performed cross-validation by testing the sample of one decoding session using a decoder trained with the samples of the remaining sessions. As a result, syllables were correctly decoded at above-chance levels. The results suggest the possibility of using non-invasively measured brain activity to read out the intended speech of disabled patients in speech motor control.