Identifying Attended Speech from Electrocorticographic Signals in a 'Cocktail Party' Setting

Dijkstra, K.V.

Identifying Attended Speech from Electrocorticographic Signals in a 'Cocktail Party' Setting

Files

Dijkstra, K.,_MA_Thesis_2014 klein.pdf (10.24 MB)

Authors

Dijkstra, K.V.

Issue Date

2014-07-01

Language

en

Abstract

People affected by severe neuro-degenerative diseases (e.g., late-stage amyotrophic lateral sclerosis (ALS) or locked-in syndrome) eventually lose all muscular control. These people are unable to use traditional assistive communication devices that depend on residual muscle control, or brain-computer interfaces (BCIs) that rely on the ability to control gaze. Auditory and tactile BCIs are considered as some of the few remaining communication options for such individuals. In this study we aimed to determine the viability of auditory attention to speech as a paradigm for BCI. We analyzed data from an experiment in which subjects at- tended to one of two speakers, to determine if the at- tended speech can be identified with better than chance performance in single trials. Our results show that we can correctly identify the at- tended speech in 7 out of 12 subjects, with an accuracy of 80% over segments of data between 4-6 seconds in length, using a regularized logistic regression. Additionally, with segments as short as 2 seconds, the average accuracy for these subjects was 70%, commonly regarded as sufficient accuracy for BCI communication. When only a single ECoG channel (i.e., cortical location) was used for classification, the attended speech could be identified in 5/12 subjects, averaging to 77% accuracy across segments 4-6 segments in length. Even though we were unable to determine why this approach failed to produce results for 5 participants, we believe that these results demonstrate the potential of this paradigm for BCI. Obvious next steps for this re- search include a further investigation of the large subject variability observed, the development of an online implementation of this paradigm and/or an expansion of the current experimental set up to determine how the obtained classification accuracy scales with an increased number of simultaneously presented speech stimuli.