VIP-ANT: VIsually-Pivoted Audio and(N) Text

May 16, 2022 ยท View on GitHub

Code for the paper Connecting the Dots between Audio and Text without Parallel Data through Visual Knowledge Transfer @ NAACL 2022.

VIP-ANT pivots audio and text via visual imagination.

Data

AudioSet can be downloaded and preprocessed via this tool.

AudioSet Data

See AudioSet. It elaborates on our customized index files for pre-training on AudioSet.

Curated Audio-Text Data

See AudioTxt. It elaborates on our curation methods and customized index files for audio-text fine-tuning.

Vision-Audio (VA) Pre-training

Check out the running script bash/run_bimodal_va.sh.

Audio-Text (AT) Fine-tuning

Check out the running script bash/run_bimodal_at.sh. Fine-tuning starts with a VA pre-trained audio encoder.

Pre-trained Models

We provide a checkpoint that performs best for each task.

Pre-trained Models for Audio-Text Retrieval

ModelAudioCapsClotho (18s)Clotho (10s)
VIP-ANT000516230004368100043681
+AT w/ GC000062100000690000004140

Pre-trained Models for Zero-shot Audio Classification

ModelESC50 (w/ prompt)US8K (w/ prompt)
VIP-ANT0008339100079420
+AT w/ GC0000414000004140

Dependencies

Dockerfile defines minimum dependencies of the repo.

License

MIT