Default Config Values

Every ESPnet3 recipe config is merged with a set of packaged defaults before it is used. You only need to write the fields you want to override — the rest come from the defaults automatically.

Where defaults come from

The defaults live in the egs3/TEMPLATE/<task>/conf/ directory inside the ESPnet3 repository. For an ASR recipe, the default files are:

egs3/TEMPLATE/asr/conf/
├── training.yaml
├── inference.yaml
├── metrics.yaml
├── publication.yaml
└── demo.yaml

These files are the canonical starting point for every ASR experiment. They contain working values for optimizer, scheduler, trainer, dataloader, and parallel settings so a new recipe does not need to redeclare everything from scratch.

How merging works: load_and_merge_config

run.py loads configs through load_and_merge_config:

training_config = load_and_merge_config(
    args.training_config,        # your conf/training.yaml
    config_name="training.yaml",
    default_package=__package__,  # egs3.TEMPLATE.asr
)

The merge order is:

1. Load the default config from egs3/TEMPLATE/asr/conf/training.yaml
2. Load your recipe config (e.g. egs3/mini_an4/asr/conf/training.yaml) without resolving interpolations yet
3. Merge — your values override the defaults, missing fields keep the default values
4. Resolve all ${...} interpolations on the merged result

This means your recipe config is always an override layer on top of the template, not a standalone file.

Tips

If a field is not in your recipe config, the value from egs3/TEMPLATE/asr/conf/ applies. You do not need to copy fields you are happy with.

Using another recipe's config as defaults

The default_package argument does not have to point to egs3.TEMPLATE. You can point it at any installed recipe package that provides a conf/ directory — for example, egs3.librispeech.asr:

training_config = load_and_merge_config(
    args.training_config,
    config_name="training.yaml",
    default_package="egs3.librispeech.asr",
)

This loads egs3/librispeech/asr/conf/training.yaml as the base and merges your recipe config on top of it.

Note

default_package can be any importable Python package path, as long as that package has a conf/ directory with the requested config file. The package must be installed or on PYTHONPATH.

Automatic inference of default_package

When default_package is not given explicitly, load_and_merge_config infers it from the path of your recipe config. A config at egs3/<recipe>/asr/conf/training.yaml maps to egs3.TEMPLATE.asr.

This means most recipes do not need to set default_package at all — the standard run.py template passes __package__ which resolves correctly for recipes that follow the egs3/<recipe>/<task>/ directory layout.

Related pages

How Configs Work

Overview of the config file layout, run.py usage, and Hydra-like patterns.

Training Config

All fields in training.yaml — model, dataset, optimizer, trainer, and more.

Config Resolvers

load_line, self_name, relative paths, and defaults composition.