amphion/Emilia-Dataset

Oct 21

There are many Cantonese pronunciations in it, and the corresponding text is wrong. Due to the large amount of data, it has a great impact on our use.
example:
ZH_B00042_S05291_W000054
ZH_B00042_S05291_W000055
ZH_B00041_S04659_W000036

laixiaofeng changed discussion title from 4/10000 AI翻译 4/10000AI translation annotation error 划译 Annotation error to data error Oct 21

yuantuo666

Amphion org Oct 21

•

edited Oct 21

Hi,
We use the whisper medium model to generate the transcript. However, the medium model did not support the yue language (check it here) which might have caused the wrong text.

To fix this, you can use a model like whisper-large-v3 or whisper-large-v3-turbo (which supports yue language) to run detect_language again, then:

filter items having mismatched language labels.
filter items with low language confidence.
filter items that have mismatched duration and text length. Check here for details.

I think these steps can help filter out those "bad" audio.

A sample:

{"id": "ZH_B00042_S05291_W000055", "wav": "ZH_B00042/ZH_B00042_S05291/mp3/ZH_B00042_S05291_W000055.mp3", "text": "的地方。", "duration": 18.364, "speaker": "ZH_B00042_S05291", "language": "zh", "dnsmos": 3.529}

HarryHe changed discussion status to closed Oct 25