ESPnet3 Config Overview

ESPnet3 uses separate YAML files per stage. Most recipes follow the layout:

egs3/<recipe>/<task>/conf/
  train.yaml
  infer.yaml
  measure.yaml
  publish.yaml
  demo.yaml

Each file is passed to run.py via the matching CLI flag:

python run.py \
  --train_config conf/train.yaml \
  --infer_config conf/infer.yaml \
  --measure_config conf/measure.yaml \
  --publish_config conf/publish.yaml \
  --demo_config conf/demo.yaml

Resolvers

ESPnet3 registers custom OmegaConf resolvers for loading external files from YAML (for example, pulling vocab or values from another config). See Resolvers for details.

Stage to config mapping

Stage	Config flag	Typical file
create_dataset	--train_config	train.yaml
collect_stats	--train_config	train.yaml
train	--train_config	train.yaml
infer	--infer_config	infer.yaml
measure	--measure_config	measure.yaml
pack_model	--train_config	train.yaml
upload_model	--publish_config	publish.yaml
pack_demo	--demo_config	demo.yaml
upload_demo	--demo_config	demo.yaml

What goes in each config

File	Purpose	Typical contents
train.yaml	Training pipeline	model, trainer, optimizers, dataloader, exp_dir
infer.yaml	Inference/decoding	model entrypoint, dataset, infer_dir, output_fn, parallel
measure.yaml	Scoring/metrics	metrics, infer_dir, test sets
publish.yaml	Packaging/upload	pack settings, artifacts to include, HF upload options
demo.yaml	Demo build	UI spec, output mapping, infer config path, assets

Notes

• Train stage: Train configuration
• Inference config: Inference configuration
• Metrics pipeline: Measurement
• Demo customization: Demo guide
• Resolvers: Resolvers