Exploring Off Time Nature for Speech Enhancement

摘要

It is well known that human conversational speech is "sparse" in time domain, comprising of many "off" time seg ments. This suggests the utility of the "off' time nature for the task of speech enhancement. We propose an efficient dual microphone method based on regularized cross-channel can cellation to distinguish the overlapping and single speech seg ments in the multi-speaker conversational environment. For tunately, the regularized cancellation results can be reused for speech enhancement along an interference-suppression chain. We present evaluations of the proposed overlapping speech de tection and integrated speech enhancement approaches using an IEEE speech database and real room recordings under various acoustic environments, showing promising improvements for speech enhancement by exploring the off time nature.