Whisper Medium 用 Monsoon 语料微调，孟加拉语词错率从 85.27% 降到 7.65%
精选理由
小模型靠数据逆袭的实测案例：Whisper Medium 微调后孟加拉语错字率从 85% 掉到 7.65%，做低资源语言 ASR 的可以参考。
@voicearena_ai 公布了其孟加拉语 ASR 语料库 Monsoon 的微调结果：769M 参数的 Whisper Medium 在 Bengali FLEURS 基准上的词错率从 85.27% 降至 7.65%。这组数据说明对低资源语言，扩充训练数据的效果优于换用更大模型。模型规模小也意味着部署成本更低，可以在靠近用户的位置运行。
原文 · rohanpaul_ai
For ASR (automatic speech recognition) on weak languages, more training data tends to beat a bigger model.
@voicearena_ai’s Bengali result for Monsoon, its ASR corpus, is shows it.
Whisper Medium fine-tuned on Monsoon went from 85.27% to 7.65% LLM word error rate on Bengali FLEURS.
Medium is the 769M-parameter Whisper. Small enough to serve cheaply, and in many settings small enough to run close to the user.
