Towards a hardware realization of time-frequency source separation of speech

This paper presents preliminary work on a hardware implementation of a source separation algorithm employing time-frequency masking methods. DUET (degenerate unmixing estimation technique) has previously been shown to achieve excellent source separation in real time in software. The current work is a move towards a hardware realization of DUET that allows integration of the algorithm into consumer devices. Initial stages involve investigating the performance of DUET when implemented in fixed-point arithmetic and a consideration of algorithmic changes to make DUET more amenable to implementation on a DSP processor. Performance is compared for floating-point and fixed-point implementations. A weighted K-means clustering algorithm is presented as an alternative to gradient descent methods for peak tracking and demonstrated to achieve excellent performance without adversely affecting computational load. Preliminary performance figures are given for an implementation on a TMS320VC5510 DSK.