Content-adaptive speech enhancement by a sparsely-activated dictionary plus low rank decomposition

摘要

One powerful approach to speech enhancement employs strong models for both speech and noise, decomposing a mixture into the most likely combination. But if the noise encountered differs significantly from the system's assumptions, performance will suffer. In previous work, we proposed a speech enhancement model that decomposes the spectrogram into sparse activation of a dictionary of target speech templates, and a low-rank background model. This makes few assumptions about the noise, and gave appealing results on small excerpts of noisy speech. However, when processing whole conversations, the foreground speech may vary in its complexity and may be unevenly distributed throughout the recording, resulting in inaccurate decompositions for some segments. In this paper, we explore an adaptive formulation of our previous model that incorporates separate side information to guide the decomposition, making it able to better process entire conversations that may exhibit large variations in the speech content.