Autonomous Augmentation of Audio Perception: Au^3

The control and communication in man and the machine has been an active area of research since the early 1940's and since then the usage of the computing machine for the augmentation or rehabilitation of man has been broadly investigated. One active area of such research is the interface of the human brain to the computer; brain-computer-interfacing (BCI). The current few successful examples of functional BCI control the computer screen cursor movement, but require extensive subject training and significant, if not full, cognitive focus. Our model proposed an alternative approach to implementing the BCI for the application of controlling a digital hearing aid by autonomously modifying the speech signal based on the identi¯cation of electrophysiological response, or an affective state. The speech communication channel was studied and the individual's hearing threshold tested while the brain activity measured using surface skin electrodes and the author's designed bioelectric amplifier. The subject acknowledgement was used as reference for which a response had previously just occurred. Using the response from a threshold 8-kHz audio tone a set of data samples is extracted for the training of a statistical learning method; the support vector machine (SVM) classifier. The SVM has the ability to produce a very high classification accuracy on small training sets of data by mapping input attribute data to span across a higher order feature space and using the features space to calculate a non-linear soft decision boundary (hypersurface) for classification. The training data output distribution was uncertain in the respect that a response had occurred, but its precise time of occurrence and duration is a component of our investigation. The SVM is taught using several output approximations and their results compared using an independent testing set of data. The reduction of data dimensionality was investigated using a Laplacian electrode array for source density estimation and the theta, alpha, beta, and gamma frequency ranges were also extracted from the Laplacian for model analysis. The results of our model were very encouraging with successful binary classification greater than 90% for raw electroencephalographic, raw Laplacian, and ¯ltered Laplacian testing measurements. Our model successfully demonstrated the e±cacy of autonomous single trial identification of affective states as an alternative or additional method of hearing prosthetic control at a communication transfer rate of 240 bits/minute.