Does whisper only support traditional Chinese？

[beiweixiaowang]

I found that the result of Chinese recognition is traditional. Do you support simplified Chinese?

[bangyuanzhu]

It may not support it, Try to use Opencc package to transfer the text.
import opencc
converter = opencc.OpenCC('t2s.json')
simplified_txt = converter.convert(txt)

[thredreams]

Or you can try zhconv

from zhconv import convert
transcribe = whisper. Decode(model, mel, options)
simplified_txt = convert(transcribe.text, 'zh-cn')

[beiweixiaowang]

Thanks to your reply, I found an interesting phenomenon, it may be
that different sizes of model training corpus are different,
whisper-base can output a part of the Chinese Simplified. I have
solved the problem by using zhconv.

def audio_to_text(model, audio_path):
    result = model.transcribe(audio_path)
    return zhconv.convert(result['text'], 'zh-cn')

[jongwook]

It may transcribe in either form since it was trained on both. Using a converter library as others suggested is a simple way to fix it, and you can also try the prompting method as shown in #277 (comment).

[beiweixiaowang]

Thanks, I'll check it out. Jong Wook Kim ***@***.***> 于2023年3月6日周一 16:46写道：

…

It may transcribe in either form since it was trained on both. Using a converter library as the others suggested is a simple way to fix it, and you can also try the prompting method as shown in #277 (comment) <#277 (comment)> . — Reply to this email directly, view it on GitHub <#987 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AOUGXLY252IQK7Y6VJY6U3LW2WP4PANCNFSM6AAAAAAVD2COXA> . You are receiving this because you authored the thread.Message ID: ***@***.***>