Summary

• The work presents a simple contrastive learning framework that doesn't require data augmentation, and only rely on sampling different segments from each sequence.
• Evaluated on nine different audio-related tasks going beyond speech, the model demonstrates superior performance over supervised counterparts.
• It conducts an extensive evaluation on different model configurations.

The problems

Prior works are limited in particular domains.

The solution

• Each input pair is sampled from a sequence, consisting two segments considered to be the positive samples. Apart from the sampling and contrastive loss, no data augmentation is used.
• Train on a large and diverse dataset, AudioSet, and evaluate on nine different audio-related downstream tasks.

Thoughts

• The simplicity may be attributed to the large dataset and the network design.
• It would be interesting to see if data augmentation degrades the performance.
• A benchmark against the reconstruction-based framework as MULTI-TASK SELF-SUPERVISED PRE-TRAINING FOR MUSIC CLASSIFICATION would be interesting.