Regarding Uyghur Dataset

#199
by osman - opened

Hi, thanks for your great work. The Uyghur dataset is working fine. However, I found out that the number of rows in the Uyghur dataset is about 21K, although CV15 includes 127 hours of validated voices. Could you please explain it? In comparison, Uzbek has 10 times more rows, although it only has 100 hours of validated voices.

Hi, @osman . Thanks for your message. That was a very interesting observation.

When I compared the Uyghur and the Uzbek datasets, I realized that, on average, the length of the samples in the Uyghur dataset was almost 10 times longer then each sample of the Uzbek one. So the Uyghur has less samples (that is, less rows) but they are longer.

Hi @fsicoli . Thanks for your quick reply. We have collected long sentences for CV. It might be the reason why Uyghur has longer raws compared to Uzbek. However, after I raised the issue here, I have checked the CV15 Uyghur dataset , and I got the following stats:

 81362 clip_durations.tsv
4688 dev.tsv
4420 invalidated.tsv
893 other.tsv
302 reported.tsv
4688 test.tsv
6529 train.tsv
76051 validated.tsv
178933 total

The Uyghur dataset includes 76051 rows in validated.tsv. However, the total Items in train, dev, and test were significantly less than in validated.tsv. Is that because of repeated recording of the same sentences? The Uzbek dataset has 140K sentences. However, the Uyghur dataset includes 20K sentences.

Hi @osman ,

To be honest, I could not find a reasonable explanation to why the train set for Uyghur is proportionally so small.

Anyway, tomorrow I will make the validated split directly available for you so you can use it in your training. In addition, I'm converting the data from Common Voice Corpus 16. That way, you will be getting some brand new data.

Hi @fsicoli thanks for your effort. I want to use CV16 once it is ready. Thanks.

osman changed discussion status to closed

Hi @osman ,

I have just released the Common Voice Corpus 16 and the Uyghur dataset is there, including the "validated" split for you. Some other languages are still being converted and loaded.

The dataset can be found at https://huggingface.co/datasets/fsicoli/common_voice_16_0

Regards,

fsicoli changed discussion status to open

Awesome, thank you.

osman changed discussion status to closed

Sign up or log in to comment