Whisper picks the wrong Chinese homophone every time: fix it with a per-model replacement table, not a bigger model

Domain: audio · Evidence E1 — Maintainer reviewed

whisperasrtranscriptionchinesehomophonepost-processing

Summary

whisper 转写中文专业内容时，同音字会**系统性**选错字，且 `initial_prompt`（领域词表）与换用更大模型（turbo→large）都修不好。实例：