SpeechBrain · SepFormer · WSJ0-2Mix
Untangle two voices.
This open demo separates an overlapping two-speaker recording into two playable audio tracks using SpeechBrain's SepFormer model.
Interactive inference is awaiting GPU activation.
The complete Gradio and ZeroGPU implementation is already in this Space repository.
The complete Gradio and ZeroGPU implementation is already in this Space repository.
The model reports 22.4 dB SI-SNR improvement on the WSJ0-2Mix test set and accepts 8 kHz mono speech mixtures.