Audio samples for the paper "One-shot conditional audio filtering of arbitrary sounds"

Authors: Beat Gfeller, Dominik Roblek, Marco Tagliasacchi

About: The audio examples in this page have been randomly selected from the evaluation split of the FSD50k dataset. Each example contains four rows: (1) clean sample; (2) mixed sample, where the clean audio has been mixed with another random example; (3) the conditioning sample; and (4) the denoised sample produced by our model. The four columns contain: (a) key information; (b) playable audio sample; (c) audio waveform; (d) mel log spectrogram for audio. In the waveform for the denoised sample, the clean audio is overlayed in orange.

We provide FSD50k examples as well as Librispeech examples (further below).

FSD50k examples


Clean sample
                  (Writing)                  
Noisy sample
(Writing + Screech)

SI-SDR=-0.01
Conditioning samples
                  (Writing)                  
Denoised sample

SI-SDR=5.48

Clean sample
                  (Applause)                  
Noisy sample
(Applause + Slam)

SI-SDR=0.06
Conditioning samples
                  (Applause)                  
Denoised sample

SI-SDR=7.55