Support character-level WER for CJK languages#101

Open

dbrkn wants to merge 2 commits into

argmaxinc:mainfrom

dbrkn:berkin/cjk-character-level-wer

Commits on Jun 8, 2026

Support character-level WER for CJK languages
dbrkn
committed
Add a dedicated CharacterErrorRate metric that splits both reference and hypothesis tokens into individual characters before alignment. This is suitable for CJK languages (Chinese, Japanese, Korean…
dbrkn
committed