Binary Mask Estimation Strategies for Constrained Imputation-Based Speech Enhancement

In recent years, speech enhancement by analysis-resynthesis has emerged as an alternative to conventional noise filtering approaches. Analysis-resynthesis replaces noisy speech with a signal that has been reconstructed from a clean speech model. It can deliver high-quality signals with no residual noise, but at the expense of losing information from the original signal that is not well-represented by the model. A recent compromise solution, called constrained resynthesis, solves this problem by only resynthesising spectro-temporal regions that are estimated to be masked by noise (conditioned on the evidence in the unmasked regions). In this paper we first extend the approach by: i) introducing multi-condition training and a deep discriminative model for the analysis stage; ii) introducing an improved resynthesis model that captures within-state cross-frequency dependencies. We then extend the previous stationary-noise evaluation by using real domestic audio noise from the CHiME2 evaluation. We compare various mask estimation strategies while varying the degree of constraint by tuning the threshold for reliable speech detection. PESQ and log-spectral distance measures show that although mask estimation remains a challenge, it is only necessary to estimate a few reliable signal regions in order to achieve performance close to that achieved with an optimal oracle mask.