LA-VocE: Low-SNR Audio-visual Speech Enhancement using Neural Vocoders
Rodrigo Mira 1 Buye Xu 2 Jacob Donley 2 Anurag Kumar 2 Stavros Petridis 1, 3 Vamsi Krishna Ithapu 1 Maja Pantic 1, 3
1 Imperial College London 2 Meta Reality Labs Research 3 Meta
Demo
(If you require the full set of evaluation samples for comparison please contact rs2517(at)ic.ac.uk)
Noise Level 1
(1 Background noise at 0 dB SNR + 1 Interfering speaker at 0 dB SIR)
comparison_with_other_works_noise_level_1.mp4
Noise Level 2
(3 Background noises at -5 dB SNR + 2 Interfering speakers at -5 dB SIR)
comparison_with_other_works_noise_level_2.mp4
Noise Level 3
(5 Background noises at -10 dB SNR + 3 Interfering speakers at -10 dB SIR)
comparison_with_other_works_noise_level_3.mp4
Spectrogram Inversion Comparison - Noise Level 2
(3 Background noises at -5 dB SNR + 2 Interfering speakers at -5 dB SIR)
spectrogram_inversion_comparison.mp4