Low Latency Audio Source Separation for Speech Enhancement in Cochlear Implants

This master thesis is a combination of two areas of Sound and Music Computing, Blind Source Separation and Cochlear Implants. The research focuses in the evaluation of existing source separation algorithms in order to improve noise reduction strategies in the context of cochlear implants. The modification and adaptation of a low latency algorithm is the point of start for the evaluation based in the requirements of speech signals in cochlear implants. The evaluation consists in a different set of objective and subjective experiments to determine the speech intelligibility enhancement produced by the separation process. Objective evaluation has revealed that a very good performance level is achieved with low latency algorithms compared to NMF which take considerably higher computation time. A series of subjective tests have been conducted with cochlear implant patients in order to compare the objective results and determine the real speech intelligibility level. The low latency algorithm showed only improvements in few situations where the noise reduction algorithm outperforms in most of the cases. Accurate analysis determined that the main reason of the speech degradation caused by low latency algorithm is because of the algorithm is not designed to detect unvoiced consonants and a lot of speech content is missing. But last experiments revealed that is possible to recover this consonants which can considerably improve the performance and later speech intelligibility of LLIS.