Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
audio
audioduration (s)
5.04
5.04

ESC-50 Fold 1 ElevenLabs Sound Effects

This dataset contains a synthetic ESC-50 fold 1 style audio set generated with the ElevenLabs Sound Effects API for an ImageBind zero-shot audio classification experiment.

It is intended for research and analysis of generated-audio domain shift. It is not an official ESC-50 dataset release.

Contents

  • audio/: 400 generated MP3 clips, 8 clips for each of the 50 ESC-50 classes.
  • meta/esc50.csv: ESC-50-compatible metadata used by the evaluation scripts.
  • generation_manifest.jsonl: generation prompts and metadata.
  • generation_plan.json: generation plan.
  • predictions_imagebind*.csv: ImageBind evaluation outputs for baseline and lightweight mitigation variants.

Evaluation Result

ImageBind huge zero-shot audio-to-text classification on this generated fold:

dataset top-1 top-5
ElevenLabs generated fold 1 55.75% (223/400) 80.50% (322/400)

For comparison, the same evaluator on official ESC-50 fold 1 produced:

dataset top-1 top-5
Real ESC-50 fold 1 65.25% (261/400) 89.50% (358/400)

Related Repository

Code and analysis reports:

https://github.com/uzumibi/report_imagebind

Notes

The audio clips were generated from ESC-50 class labels with prompt and prompt_influence variation to create class-preserving diversity. Please check ElevenLabs terms and policies for permitted use of generated sound effects.

Downloads last month
46