Datasets:
samples of another language are mixed in a webdataset
Hi, thanks for the release of this amazing dataset!
It really helps the TTS community. Truly appreciate your effort.
I have one inquiry regarding a language mix in the data.
In a .tar where it is tagged as for example Korean, there are several English samples within. For example,
- Mandarin tar in Emilia contains several English data
- Korean tar in Emilia_yodas contains several English data
As far as I understood, you set whisper.detect_language() prob threshold as a certain value to classify the languages.
So this might be the problem of whisper?
Since this is quite a critical problem, I would like to kindly ask you to check this issue out.
Or, could you provide a good way to handle this out in my end?
e.g., set prob threshold higher and filter out the ambiguous samples?
Thanks a lot in advance!
Hope you have a great day :)
Hi, @sjchung
Thank you so much for your attention to our work. Processing in-the-wild speech data for training is indeed a challenging task. While we are leveraging almost all state-of-the-art components in the Emilia pipeline, there are inevitably edge cases that arise (e.g., mixed languages, multiple speakers in a single utterance). Although our experiments have demonstrated the effectiveness of the current pipeline, we are optimistic that future advancements in speech technology will help alleviate these challenges.
For your specific use case, I recommend using the Emilia-Yodas subset, where the source data is derived from the Yodas dataset, complete with language annotations. This should result in less language-mixed data for your needs. Additionally, as you suggested, setting a higher probability threshold and filtering out ambiguous samples can certainly help (and I can confirm that language detection using whisper.detect_language() is highly efficient). However, please note that this approach may involve a trade-off, as it could lead to the exclusion of some correct samples.
I hope this response helps address your inquiry.