Finetuning

Most ESPnet3 finetuning work reduces to three questions:

1. Which dataset do you train on?
2. Which model do you actually fine-tune?
3. How much of the old model do you keep or replace?

Use this section for those three topics.

Custom dataset

Point training and inference at your own dataset module and manifests.

Customize the model

Use `src/model.py`, keep or replace old weights, and switch from the task bridge when needed.

Adding a stage

Use the contributor guide when fine-tuning needs extra preparation or export stages.

Customize the training loop

Use recipe-local LightningModule, trainer, or System code when the default loop is not enough.