Vowel classification using wavelet decomposition during speech imagery

摘要

Electroencephalography (EEG) has long been used for Brain computer interface (BCI). Recent researches have proved that EEG can be also used to classify data generated in speech imagery. This classification can further be utilized to develop speech prosthesis and synthetic telepathy systems. In this paper we wanted to check whether features extracted from beta, delta and theta rhythms of EEG can be used to classify the imagined English vowel sounds. A new approach is used to differentiate among the three classes of vowel sound /a/, /u/ and `rest or no action' in pair-wise as well as `combination of two sounds (tasks)' manner. Wavelet decomposition is performed to extract features in the 0-8 Hz and 16-32 Hz range. Energy sum and energy's waveform length of the approximate and detail coefficients are used as features. The algorithm is tested on 3 subjects and results showed that indeed the data from EEG rhythms can be used for classification. The pair-wise classification accuracy was found to be 65-82.5% which is a considerable improvement over the previous classification accuracies in the range of 56-82%, reported by DaSalla [4]. The `combination of tasks' classification accuracy was found to be 81.25-98.75%.