amphion/Emilia-Dataset · Missing metadata & multilingual problems

Sep 6

First I would love to thank you for ur solid work and contributions.
When I dealing with the dataset, I ran into some small issues:

The metadata (Text) for DE_B00000_S06595_W000024.mp3, whose ASR result is "Das Feuer ist erloschen. Sind alle in Sicherheit?" , is missing. I fixed it myself but I think someone can update it if the missing problem does exist.
Also I find some data may contain multilingual audios, e.g.,DE_B00000_S07871_W000002, which the audio is

好的，我今天已经可以开始健身了吗，我今天就开始吗 ？Ja klar, zuerst machen wir einen Gesundheitscheck und besprechen, was du machen möchtest.

but the text metadata is

11.817( duration ) 3.0946 (dnsmos) Ja klar, zuerst machen wir einen Gesundheitscheck und besprechen, was du machen möchtest.
so I wonder if anyone have any idea how to filter out these data. I myself first tried whisper which is of course not work (because it is the same one used to create the dataset), and currently my solution is to compare the usual tts time considering text length and filter out abnormal ones, and I also tried to use "langid" to filter out some English&German sentences(e.g DE_B00000_S07875_W000000 ). I hope this can be of some help.
if someone have better solution, let me know pls.

Amphion org Sep 6

Hi, thanks for your usage and feedback. Here is some response to your comments:

We intentionally deleted some metadata (the total should be around several hours) since they are either low DNSMOS or some hallucination occurred in the transcribed text. We detect these by using the DNSMOS value and text-based repeat detection.
For these multilingual audios, we can detect them by using the prob return value when calling whisper.detect_language() and, as you mentioned, calculate the duration per character. We already implemented these two in our pipeline. You might change them and set a higher bar to lower the probability of having multiple languages in the same segment.

HarryHe changed discussion status to closed 28 days ago