Reproduce “Weakly-supervised Automated Audio Captioning via text only training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Theodoros Kouzelis, Vassilis Katsouros
Researchers from Athena Research Center developed a weakly-supervised method for Automated Audio Captioning (AAC) that trains models using only text, significantly reducing reliance on paired audio-caption data while achieving performance comparable to fully supervised approaches.
Key insights
- CLAP models, similar to CLIP, exhibit a 'modality gap' where semantically similar audio and text embeddings occupy distinct regions in the shared embedding space.
- Both training and inference-time strategies are effective in mitigating this modality gap, significantly boosting captioning performance over the baseline text-only training.
- Inference-time strategies, particularly Projection-Based Decoding, generally outperform training-time strategies, likely due to their ability to leverage a 'Memory' set of in-domain text embeddings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.