From explicit to implicit speech recognition

Abstract

We consider the problem of determining the word or concept that a subject holds in their mind prior to the act of speech using only a scalp-recorded electroencephalogram (EEG). Such speech acts are called covert, silent, or implicit speech acts in the literature. We consider a binary-tree classifier that uses one of a number of candidate feature types, including temporal correlation coefficients, spectral correlation, and time-gated raw voltages. The particular features and binary-tree parameters are blindly determined using the local discriminant basis (LDB) technique. The experiments involve sequential presentation of words and numbers on a computer screen. The subject wears an EEG scalp cap and is instructed to first consider the stimulus, then speak it. Later, the subject is instructed to perform the same task without the actual utterance, resulting in implicit speech. We present performance results for the various obtained classifiers, which show that the approach has significant merit. © 2013 Springer-Verlag Berlin Heidelberg.