Transcription quality?

#2
by cckm - opened

Thank you for contributing this dataset!

Just wondering if you have some statistics or general feelings on the accuracy of the transcriptions?

@cckm
Oh, sorry for the delay.

I didn't do any EDA, so I don't have any results to share with you.

I know that KoreaSpeech doesn't have the best transcription quality, especially for the speech data I uploaded. All I can recommend is to tighten up the processing.

Thanks for your reply! I spot checked a few entries, and it seems the transcription is consistent with the output of whisper large v3 turbo (https://huggingface.co/openai/whisper-large-v3-turbo). whisper large v3 turbo is pretty good for Korean (https://github.com/openai/whisper/discussions/2363).

I may try to use this for TTS, after filtering out "quality" != "정상", and transcriptions that contain alternatives (like "(감자)/(감짜)").

Thanks for the advice! I'll keep it in mind for future data preprocessing.

jp1924 changed discussion status to closed

Sign up or log in to comment