Optimal Condition Training (OCT) for Target Source Separation

Code and data recipes for the paper: Optimal Condition Training for Target Source Separation by Efthymios Tzinis, Gordon Wichern, Paris Smaragdis and Jonathan Le Roux

TLDR; The main contribution of this paper is to introduce a novel way of training, namely, optimal condition training (OCT) for single-channel target source separation, based on greedy parameter updates using the highest performing condition among equivalent conditions associated with a given target source. OCT improves upon single-conditioned models and oracle permutation invariant training. We also propose a variation of OCT with condition refinement, in which an initial conditional vector is adapted to the given mixture and transformed to a more amenable representation for target source extraction.

Table of contents

• Paths Configurations
• How to use the dataset loaders
• References

Paths Configurations

Change the dataset paths to the ones stored locally for all the root directories (the metadata paths are going to be created after runnign the scripts presented below):

git clone https://github.com/etzinis/optimal_condition_training.git
export PYTHONPATH={the path that you stored the github repo}:$PYTHONPATH
cd optimal_condition_training
vim __config__.py

You should change the following:

ROOT_DIRPATH = "the path that you stored the github repo"
FSD50K_ROOT_PATH = "the path that you stored FSD50K"

How to use the dataset loaders

Now that everything is in place, one might use the combined dataset loader which enables to create the three different versions of FSD50K [1] dataset.

• Random super-classes: We first randomly sample two distinct sound classes (out of the available 200), then sample a representative source waveform for each class and mix them together. is_hierarchical=False, intra_class_prior=0.
• Different super-classes: We select a subset of classes from the FSD50K ontology corresponding to six diverse and more challenging to separate super-classes of sounds, namely: Animal (21 subclasses), Musical Instrument (35 subclasses), Vehicle (15 subclasses), Domestic & Home Sounds (26 subclasses), Speech (5 subclasses) and Water Sounds (6 subclasses). Each mixture contains two sound waveforms that belong to distinct super-classes. is_hierarchical=True, intra_class_prior=0.
• Same super-class: Following the super-class definition from above, we force each mixture to consist of sources that belong to the same abstract category of sounds to test the ability of text-conditioned models in extremely challenging scenarios. is_hierarchical=True, intra_class_prior=1.

By channging the following variables we can also control the samplign probability of each heterogeneous condition [2]:

valid_conditions = ["harmonicity", "energy", "source_order",
                    "one_hot_class", "one_hot_super_class", "multi_hot_class"]
queries_priors = [0.4] + [0.3] + [0.3] + [0.] * 3

The dataset uses online mixing. You can test the generator as shown next:

cd optimal_condition_training/dataset_loader 
➜  dataset_loader git:(main) ✗ python heterogen_fsd50k_2mix.py
For randomly sampling between 200 classes of sounds in FSD50K.
Sampled the classes: ('Dishes', 'Crushing', 'Cat', 'Telephone', 'Doorbell', 'Sigh', 'Singing', 'Chirp')
Sampled the following conditional vectors: tensor([[0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 1., 0.,  ..., 0., 0., 0.],
        ...,
        [0., 1., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.]])
It took me: 1.7711923122406006 secs to fetch the batch

For a hierarchical FSD50K version with probability of sampling from the same super class: 0.0
Sampled the classes: ('Stream', 'Bathtub', 'Child speech', 'Writing', 'Toilet flush', 'Frog', 'Child speech', 'Livestock')
Sampled the following conditional vectors: tensor([[0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [1., 0., 0.,  ..., 0., 0., 0.],
        ...,
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 1.,  ..., 0., 0., 0.]])
It took me: 0.8072104454040527 secs to fetch the batch

For a hierarchical FSD50K version with probability of sampling from the same super class: 1.0
Sampled the classes: ('Child speech', 'Tambourine', 'Boat', 'Drum kit', 'Truck', 'Waves', 'Wind chime', 'Stream')
Sampled the following conditional vectors: tensor([[1., 0., 0.,  ..., 0., 0., 0.],
        [1., 0., 0.,  ..., 0., 0., 0.],
        [1., 0., 0.,  ..., 0., 0., 0.],
        ...,
        [0., 0., 0.,  ..., 0., 0., 0.],
        [1., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.]])
It took me: 1.0530891418457031 secs to fetch the batch

References

[1] E. Fonseca, X. Favory, J. Pons, F. Font, and X. Serra, “Fsd50k: an open dataset of human-labeled sound events,” arXiv preprint arXiv:2010.00475, 2020.

[2] Tzinis, E., Wichern G., Subramanian, A., Smaragdis, P., and Le Roux, J., “Heterogeneous target speech separation.” In Proceedings of Interspeech, 2022, pp. 1796-1800.

[3] Tzinis, E., Wang, Z., Jiang, X., and Smaragdis, P., “Compute and memory efficient universal sound source separation.” In Journal of Signal Processing Systems, vol. 9, no. 2, pp. 245–259, 2022, Springer