Skip to main content
Back

Google DeepMind's SURF Separates Audio Without Any Clean Training Data

SURF trains a generative flow-matching model to separate audio and images using only raw mixtures, with no clean source data required, setting a new unsupervised state-of-the-art.

Paper, Full Access
  • SURF is the first generative source separation model that trains entirely from raw mixtures, with no clean source data required.
  • It combines flow matching with a teacher-student remixing loop: a teacher separates mixtures, estimates are shuffled and recombined, and a student flow model trains on these synthetic pairs.
  • On MNIST image separation, SURF (ReMixIT) achieves PSNR 37.26 vs. 23.13 for the best prior unsupervised method, nearly matching supervised flow matching.
  • On CIFAR-10, SURF achieves FID 12.50, outperforming BASIS (FID 22.12) which uses a supervised diffusion prior trained on clean images.
  • On Libri2Mix speech separation, SURF reaches 16.28–16.54 dB SI-SDR vs. 12.39–13.60 dB for prior unsupervised methods, approaching the supervised ceiling of 17.89 dB.
  • The paper draws a novel theoretical connection to the Wake-Sleep algorithm and was accepted to ICML; a demo page with audio examples is available.

Every time you've used a noise-canceling headset or watched a movie with clean dialogue pulled from a noisy set recording, source separation was doing the heavy lifting. The field has made enormous strides with supervised deep learning, but there's a stubborn bottleneck: you need large libraries of clean, isolated recordings to train these models. For many real-world domains, that data simply doesn't exist. A new paper from Google and Google DeepMind, SURF: Separation via Unsupervised Remixing Flow, proposes a way around this wall entirely.

The clean-data problem

Single-channel source separation is the task of taking one recording that contains multiple overlapping signals and recovering each individual source. Think of unmixing two speakers talking at once, or separating a voice from background noise. The challenge is deeply ill-posed: infinitely many combinations of sources could sum to the same mixture.

Modern generative approaches, particularly diffusion and flow-matching models, have tackled this by learning a strong prior over what clean signals look like, then using that prior to guide separation. The catch is that training these priors requires access to large datasets of clean, isolated sources. In many complex domains such as bio-acoustics, hyperspectral imaging, or gravitational wave detection, obtaining such clean, in-domain source data is practically impossible. And even when clean data exists, supervised models are vulnerable to domain shifts.

Prior unsupervised approaches like MixIT and ReMixIT addressed the data problem for regression-based models, but no one had successfully applied this idea to generative flow-matching models. That's the gap SURF fills.

Flow matching, briefly

Flow matching (FM) is a generative modeling technique that learns to transport samples from a simple noise distribution to a target data distribution by training a neural network to predict a "velocity field" , essentially, the direction and speed to move a noisy sample at each point in time to reach a clean sample. For source separation, the model starts from a noisy version of the mixture and flows toward the separated sources.

The key difficulty in going unsupervised is that flow matching normally requires pairs of (noisy, clean) training examples. SURF's core contribution is a way to generate those pairs synthetically, using only the raw mixtures themselves.