The audio tracks from the synced video clips are fed into our algorithm, that produces an enhanced audio (1st column).
        All Youtube clips are with different YoutubeID.

Example Ⅰ

Real world data, Harry Styles concert.

our crowdsourced enhancement result video (1) video (2) video (3) video (4)
Note the clear scream pattern in the spectrogram of video (2) at 1.5 seconds, that is not present in the enhanced result.

Example Ⅱ

Real world data, Coldplay concert.

our crowdsourced enhancement result video (1) video (2) video (3) video (4)
In the enhanced audio the applause heared in videos (1) and (2) are significantly reduced, and the scream from video (4) is absent.

Example Ⅲ

Real world data, Harry Styles concert (2).

our crowdsourced enhancement result video (1) video (2) video (3) video (4) video (5)
video (6) video (7) video (8) video (9)


Example Ⅳ

Synthetic data, speech + noises (both sampled from LibriSpeech), SNR=-3dB.

our crowdsourced enhancement result signal + noise (1) signal + noise (2) signal + noise (3) signal + noise (4) signal + noise (5)


Example Ⅴ

Synthetic data, speech (from LibriSpeech) + noises sampled from Audioset, SNR=3dB.

our crowdsourced enhancement result signal + noise (1) signal + noise (2) signal + noise (3) signal + noise (4) signal + noise (5)


Example Ⅵ

Synthetic data, pyroomacoustics room simulation, speech and noises from LibriSpeech, SNR=-3dB.

our crowdsourced enhancement result signal + noise (1) signal + noise (2) signal + noise (3) signal + noise (4) signal + noise (5) signal + noise (6) signal + noise (7)


Example Ⅶ

Synthetic data, pyroomacoustics room simulation, speech and noises from LibriSpeech, SNR=0dB.

our crowdsourced enhancement result signal + noise (1) signal + noise (2) signal + noise (3) signal + noise (4) signal + noise (5) signal + noise (6) signal + noise (7)