- CLIPSonic: text-to-audio synthesis with unlabelled videos.
Our models trained without text-audio pairs, can generate audio from text. - Adversarial PIT for universal sound separation
Listen how our adversarial setup reduces the spectral holes when separating. - Full-band general audio synthesis with score-based diffusion
Check our examples of audio synthesis and style transfer. - Upsampling artifacts in neural audio synthesis
Check our site summarising our main findings (with examples, code and a video). - End-to-end music source separation: is it possible in the waveform domain?
Read our paper, and listen to some separations. - A Wavenet for speech denoising
Read our paper, and listen to some examples. - Musically motivated CNNs for music auto-tagging
Read our paper, and listen to some music examples.
Check out musicnn, our open-source music tagging system.
Webpages of datasets I helped creating:

That’s me demonstrating I had a guitar and a cool camera.