游雁 bd9eeabf14 feat: implement CR-CTC (Consistency-Regularized CTC) loss ...

CR-CTC improves CTC training by computing KL divergence between encoder
outputs from augmented and clean inputs. This enforces prediction consistency
regardless of data augmentation.

k2/icefall reports 3.98% CER on AISHELL-1 test with Zipformer + CR-CTC.

Usage: Add model_conf.cr_ctc_weight=0.2 to enable CR-CTC during training.
The clean encoder pass is done without SpecAug while the augmented pass
uses the normal SpecAug pipeline.

Also fixes: add_sos_and_eos function for EParaformer compatibility.

2026-05-25 22:45:14 +08:00

..

__init__.py

update details

2022-11-29 12:57:21 +08:00

cr_ctc.py

feat: implement CR-CTC (Consistency-Regularized CTC) loss

2026-05-25 22:45:14 +08:00

label_smoothing_loss.py

docs: add docstrings to 1950 functions across 276 files for API documentation

2026-05-21 07:52:52 +00:00