Supervised single-channel speech enhancement using ratio mask with joint dictionary learning

More comparisons and analyses about the paper "Supervised Single-Channel Speech Enhancement Using Ratio Mask with Joint Dictionary Learning" mentioned are given in this document. Specifically, 1) comparisons between the clean spectra/IRM and reconstructed spectra/IRM in the training stage are done; 2) the influence of the speech dominant parameter  performed on the SNMF algorithms are shown; 3) a Wiener-type filter and the proposed SM filters applied to GDL algorithm are performed and discussed; 4) discussions about different ways using the estimated RMs to recover the speech from the mixture are presented; 5) some more results of another 4 different noise types for the speaker-dependent case are shown. The previous four experiments have been done under the speaker-dependent case. 6) The more results for the speaker-independent case are shown to compare the performances of different algorithms. At last, the spectrograms of used interferers are given at last. As mentioned in our paper that the 3 objective measures (segSNR, PESQ, fwsegSNR) are calculated to measure the objective intelligibility and quality of recovered speech from different perspectives, although there are some correlations between them (Ma, 2009; Hu, 2008). The objective measures may be improved to various extent by different algorithms. From this perspective, the improvements of segSNR, PESQ and fwsegSNR may not be consistent between RMJDL-SM1 and RMJDL-SM2, which means one algorithm may improve one measure more than another algorithm, but the situation may be reverse for another measure.