Datasets:

amphion
/

Emilia-Dataset

Hi there, thanks for sharing this awesome dataset!
I am trying to process and combine the remaining (unused) YODAS data with tis Emilia Large dataset, could you please share the unused data list if it exists? Also, it will be great if you can explain the rationale (e.g., language, DNSMOS score) how the Emlia-YODAS is selected from the YODAS dataset, as the original YODAS dataset is around 422k hours and Emilia-YODAS is only ~114k hours?

HarryHe

Amphion org Aug 18

Hi @zxsu , Thank you for your interest in our dataset and question. Please refer to Sec.~IV.B in our paper for these details.

zxsu

Aug 18

•

edited Aug 18

Hi @HarryHe , thanks for your reply! Does it mean that all data in YODAS that has a DNSMOS score above 2.4 is included in Emilia-YODAS?

HarryHe

Amphion org Aug 19

Hi @HarryHe , thanks for your reply! Does it mean that all data in YODAS that has a DNSMOS score above 2.4 is included in Emilia-YODAS?

Yes, all (en, zh, jp, kr, de, fr) data in YODAS that has a DNSMOS score above 2.4 is included in Emilia-YODAS. Data in other languages are not included.

zxsu

Aug 19

Thanks again, @HarryHe this information is very helpful

zxsu

Aug 23

•

edited Aug 23

Hi @HarryHe , sorry for another question – I noticed that most (if not all) audios in Emilia-YODAS look like upsampled from 16kHZ (8k-12k band is black in spectrogram). Did you use the yodas2 dataset in 24kHz or the yodas dataset (which is segmented and resampled to 16kHz)?

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment