Audio Demos for 2-Source Universal Sound Separation


Ice water + telephone clunks


Mixture
Ground-truth sources

SI-SDR = 0.97 dB


SI-SDR = -1.06 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = 16.97 dB


SI-SDR = 15.94 dB
iTDCN++ 2.5ms learned

SI-SDR = 15.38 dB


SI-SDR = 14.29 dB

Morse code + crash


Mixture
Ground-truth sources

SI-SDR = -0.67 dB


SI-SDR = 0.70 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = 11.81 dB


SI-SDR = 12.52 dB
iTDCN++ 2.5ms learned

SI-SDR = 5.52 dB


SI-SDR = 7.31 dB

Dog bark + shaving foam


Mixture
Ground-truth sources

SI-SDR = -0.79 dB


SI-SDR = 0.79 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = 10.57 dB


SI-SDR = 11.45 dB
iTDCN++ 2.5ms learned

SI-SDR = 9.87 dB


SI-SDR = 10.80 dB

Golf ball + basketball shoot


Mixture
Ground-truth sources

SI-SDR = -15.60 dB


SI-SDR = 14.80 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = 2.70 dB


SI-SDR = 19.17 dB
iTDCN++ 2.5ms learned

SI-SDR = 4.27 dB


SI-SDR = 20.20 dB

Seagull squawking + car drop


Mixture
Ground-truth sources

SI-SDR = -18.98 dB


SI-SDR = 18.67 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = -11.38 dB


SI-SDR = 18.89 dB
iTDCN++ 2.5ms learned

SI-SDR = -15.06 dB


SI-SDR = 17.90 dB

Operator need help + metal door


Mixture
Ground-truth sources

SI-SDR = 8.15 dB


SI-SDR = -8.13 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = 14.37 dB


SI-SDR = 5.22 dB
iTDCN++ 2.5ms learned

SI-SDR = 14.29 dB


SI-SDR = 5.15 dB

Metal widget + insect voice scream


Mixture
Ground-truth sources

SI-SDR = 42.65 dB


SI-SDR = -43.58 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = 8.64 dB


SI-SDR = -40.37 dB
iTDCN++ 2.5ms learned

SI-SDR = 7.55 dB


SI-SDR = -41.18 dB

Footsteps wood stairs + refrigerator water dispenser


Mixture
Ground-truth sources

SI-SDR = -15.40 dB


SI-SDR = 15.58 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = 2.24 dB


SI-SDR = 19.85 dB
iTDCN++ 2.5ms learned

SI-SDR = -2.61 dB


SI-SDR = 16.20 dB

Automatic door close + car emergency brake


Mixture
Ground-truth sources

SI-SDR = 13.49 dB


SI-SDR = -12.40 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = 14.50 dB


SI-SDR = -4.49 dB
iTDCN++ 2.5ms learned

SI-SDR = 15.24 dB


SI-SDR = -2.34 dB

Metal hit + car wiper


Mixture
Ground-truth sources

SI-SDR = -5.00 dB


SI-SDR = 5.13 dB
Method Separated source 0 Separated source 1
iTDCN++ 2.5ms STFT

SI-SDR = 3.58 dB


SI-SDR = 9.67 dB
iTDCN++ 2.5ms learned

SI-SDR = 4.72 dB


SI-SDR = 10.32 dB