A real-time implementation of constrained estimate-maximize algorithm for single-microphone speech enhancement

A real-time frequency-domain implementation of constrained estimate-maximize (EM) algorithm is presented to work on a single-microphone system. Our experiments based on a moderate speech database indicate that the performance of the proposed implementation is comparable with an off-line approach subject to a matching constraint. A DSP chip operated above 33 MHz is competent to carry out 6 iterations of the proposed algorithm with a delay time of 16 ms.