External Open-Music Ingestion

Goal

Convert local open-music audio folders into ACR-ready manifests for:

• training queries
• evaluation queries
• reference catalog indexing

Recommended personal-use flow

1. Prepare a local audio directory

Examples:

• data/raw/fma_small_audio/
• data/raw/mtg_jamendo_audio/

Drop-zone details:

• data/raw/README.md

2. Generate manifests through the adapter entrypoint

Optional pre-check:

/usr/local/miniconda3/bin/python src/data/external_adapters.py inspect-local fma data/raw/fma_small_audio --eval-ratio 0.2 --query-duration 8.0

Batch pre-check across multiple candidate corpora:

/usr/local/miniconda3/bin/python src/data/external_adapters.py inspect-batch fma=data/raw/fma_small_audio mtg_jamendo=data/raw/mtg_jamendo_audio --eval-ratio 0.2 --query-duration 8.0

Then generate manifests:

/usr/local/miniconda3/bin/python src/data/external_adapters.py prepare-local fma data/raw/fma_small_audio --output-root data/external_ingested --eval-ratio 0.2 --query-duration 8.0

or

/usr/local/miniconda3/bin/python src/data/external_adapters.py prepare-local mtg_jamendo data/raw/mtg_jamendo_audio --output-root data/external_ingested --eval-ratio 0.2 --query-duration 8.0

3. Use outputs

Generated files:

• catalog.json: reference tracks for indexing
• train.json: train queries + references
• test.json: held-out eval queries + references
• val.json: optional validation split

Notes

• Small datasets are automatically protected so both train/test query sets exist.
• For personal use, FMA and MTG-Jamendo should be the first real baselines.
• Keep test.json fixed across experiments to compare models fairly.