Comparison of Parameter Estimation Methods for Single-Microphone Multi -Frame Wiener Filtering

The multi-frame Wiener filter (MFWF) for single-microphone speech enhancement is able to exploit speech correlation across consecutive time-frames in the short-time Fourier transform (STFT) domain. To achieve a high speech correlation, typically an STFT with a high time-resolution but a low frequency-resolution is applied. The MFWF can be decomposed into a multi-frame minimum power distortionless response (MFMPDR) filter and a single-frame Wiener postfilter. To implement the MFWF using this decomposition, estimates of several parameters are required, namely the speech correlation vector, the noisy speech correlation matrix, and the power spectral densities at the output of the MFMPDR filter. Correlations can be estimated either directly in the low frequency-resolution STFT filterbank, indirectly by estimating periodograms in a high frequency-resolution filterbank and applying the Wiener-Khinchin theorem, or in a combined way. In this paper, we compare the performance of different estimators for the required parameters. Experimental results for different speech material, noise conditions, and signal-to-noise ratios show that using a combined estimator for the speech correlation vector yields the best results in terms of speech quality compared to existing direct and indirect estimators.