Augmenting information channels to improve cochlear implant patients performance under adverse conditions

Cochlear Implant patients perform reasonably well in acoustically pristine environments. However, performance degrades significantly under adverse conditions, particularly within speech-like noisy surroundings. This thesis addresses the problem of resolving starts and ends of spoken words to improve speech intelligibility by CI patients under severe adverse conditions.; We propose a new approach based on sparse representation of speech signals. The approach is based on the discrete wavelet packet decomposition stemming from its excellent ability to capture transient signals. The obtained sparse representation is parameterized using a Gaussian mixture model to yield a feature set for classification and clustering purposes. We compare two methods to classify the start and end segments of spoken words in a noisy environment. The first relies on exploiting second order statistics of the sparsely represented signals, while the second relies on an expectation maximization approach to the Gaussian Mixture Model. We test the performance of both methods under various signal and noise conditions. Our preliminary results demonstrate that the sparse representation can capture eminent features in the spoken words that are indicative of start and end of the word. The proposed approach can be useful in driving cochlear implant signal transduction mechanisms with new features that are more robust to adverse conditions than the classical filter bank approach commonly used in current technology.