After finishing the paper, we found that the buzzing artifacts of unvoiced and semi-voiced segments are from the harmonic part signals, and only occur in the ones based on subtractive synthesizers (i.e. in model SawSinSub, SawSub and Full). To remove them, we apply a Voiced/Unvoiced mask on the harmonic part singals. The mask is currently estimated from the predicted signals, in the future this could be integrated as a part of the model, or given by acoustic models. Note that this step is excluded from the user study in our paper.
- Run
debuzz.ipynb - In the
test-setfolder:debuzz: post-processed de-buzzing resultspart: separated harmonic and noise part signalspred: orignal predicted signals
Special thanks to Oscar Friedman.