ESPnet3 Recipe Directory Layout

Current recipes under egs3/ typically look like:

egs3/<recipe>/<task>/
  run.py
  conf/
    training.yaml
    inference.yaml
    metrics.yaml
    publication.yaml
  dataset/
    __init__.py
    builder.py
    dataset.py
  src/
    inference.py
    tokenizer.py
  data/
  exp/

Where to put what

Location	You put here	Typical contents
egs3/<recipe>/<task>/conf/	YAML configs	training, inference, metrics, publication
egs3/<recipe>/<task>/dataset/	dataset module	Dataset, DatasetBuilder, recipe-local dataset exports
egs3/<recipe>/<task>/src/	custom Python helpers	inference helpers, tokenizer logic, recipe-local system code
egs3/<recipe>/<task>/run.py	entry script	stage list, config loading, system selection

Important directories

• conf/: stage configs
• dataset/: recipe-local Dataset and DatasetBuilder
• src/: extra recipe code such as output functions or tokenizer helpers
• data/: prepared manifests or recipe-local artifacts
• exp/: checkpoints and experiment outputs

run.py

run.py loads the stage configs, applies naming propagation, instantiates the system, and runs the selected stages.

Current config flags are:

• --training_config
• --inference_config
• --metrics_config
• --publication_config

Typical stage lists include:

• create_dataset
• collect_stats
• train
• infer
• measure
• pack_model
• upload_model

Directory structure

More broadly, recipes under egs3/ follow a shared pattern:

egs3/
  TEMPLATE/
    <task>/
      run.py
      conf/
  <recipe>/
    <task>/
      run.py
      conf/
      dataset/
      src/
      data/
      exp/

TEMPLATE provides the default config values and stage runner wiring, while each concrete recipe overrides only what it needs.

Creating a new recipe

1. Copy egs3/TEMPLATE/<task> into egs3/<your_recipe>/<task>.
2. Add configs under conf/.
3. Implement recipe-local dataset code under dataset/ when needed.
4. Add src/ helpers such as inference.py or recipe-local system code only when the recipe needs them.
5. Run stages through python run.py --stages ....

Notes

The main dataset story is now dataset/, not src/dataset.py as the primary integration point.