Separation of Periodic and Stochastic Components in Real-Time Sound Devices for Improved Representation in Speech Enhancement

In traditional real-time speech enhancement a fixed filter bank is utilized to process individual frequency bands, which put restrictions on the time/frequency resolution. To accommodate more advanced speech enhancement techniques it is desirable to simultaneously increase resolution in time and frequency. In speech coding a solution to this problem is to separate periodic and stochastic components to achieve an improved representation for each component. The goal of the project is to replace the generic processing scheme in real time speech enhancement with a scheme for separation of signal components that are deterministic from components that are stochastic, and to develop a dual processing scheme consisting of systems tailored for processing each set of components. This must occur quickly (within a few milliseconds) and perform reliably in poor signal-to-noise ratios as well as obey constraints on complexity and delay for real-time sound devices such as hearing aids. Also the project will develop multi-microphone algorithms that use spatial information to improve the separation.