ESPnet2 To ESPnet3 Migration

The main migration shift is:

• ESPnet2: shell-driven recipes and stage numbers
• ESPnet3: Python run.py, System methods, YAML-first configs

Current ESPnet3 naming

Use:

• training.yaml
• inference.yaml
• metrics.yaml
• publication.yaml

and stages:

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

Dataset migration

Recipe-local dataset integration should now center on:

• dataset/__init__.py
• dataset/builder.py
• dataset/dataset.py

not on the older src/create_dataset.py / src/dataset.py framing as the main integration story.

Practical migration checklist

1. copy egs3/TEMPLATE/<system>/
2. move config naming to training.yaml, inference.yaml, metrics.yaml, publication.yaml
3. port data preparation into DatasetBuilder
4. port dataset loading into Dataset
5. update run.py to use current config flags
6. update inference and measurement to infer + measure

Recommended references

• Recipe directory layout
• Dataset references and builders
• Systems