Mixture Invariant Training (MixIT)

Definition

Training paradigm for source separation that uses mixtures of mixtures (MoMs) as input instead of requiring isolated ground-truth sources. The model separates a MoM into its constituent mixtures, which are compared against the known sub-mixtures.

Key Ideas

• Wisdom, Hershey et al. (Google). Take two mixtures, sum them = MoM. Model separates MoM into estimates of the original mixtures.
• Loss: compare estimated outputs to original mixtures (not to isolated sources).
• Key advantage: does not require isolated stems for training. Can use any audio.
• Works because the mapping from MoM to constituent mixtures is unambiguous (no permutation problem).
• Enables training on in-the-wild data at massive scale.

Relationships

• From john-hershey's group at Google
• Related to ../concepts/permutation-invariant-training, ../concepts/deep-clustering-separation
• Contrast with ../concepts/synthetic-mixing-pipelines — MixIT doesn't need isolated stems
• Implemented in some ../entities/asteroid recipes

Sources

None ingested yet — seed batch setup.