chenchen

How to get correct word from bpe subword

2

Hello， When I use bpe as the modeling unit to train the English ASR model, the output of model are bpe subwords, and words can be obtained by spaces and...

whisper模型微调后转为faster-whisper模型，流式解码的时候效果变差

3

您好， 我最近用wenet里的whisper-finetune代码微调了一个whisper模型，然后将他转为faster-whisper支持的模型，用whisper streaming的代码进行流式解码，发现微调后的模型解码效果变差了很多（存在比较多的漏字问题），用wenet解码是正常的 。请问下会是什么原因呢？

The inference speed becomes slower when I use the fine-tuned model and set word_timestamp to True

1

Hi, When I finetune the whisper-large-v3 model， I use convert_openai_to_hf.py to convert model to hf model, then use ct2-transformers-converter . However, If I set word_timestamp = True, The decoding speed...