Hugging Face's logo

Datasets:
voice-is-cool
/
voxtube

Tasks:
Audio Classification
Modalities:
Audio
Text
Formats:
parquet
Languages:
English
Russian
Spanish
Size:
1M - 10M
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files
xet
 Community
2

Thank you. Are these dataset overlapped ?

#2
by thusinh1969 - opened
Discussion
thusinh1969

Thanks for accepting me.
I have a few questions please:

  1. Is speaker ID checked that is it unique or it can be overlapped with other channels same persons --> can not use for SV
  2. One YT movie may have more than one person speaking (podcast etc...), do we actually scan them out

At the end, uniqueness if what I am after as it may be a hit.miss for training speaker verification model.

Thanks much for your answer.
Steve
antonioperfecto
 voice-is-cool org

Hi Steve!

  1. Speakers ids aren't overlapped - i.e. duplicated speakers with high similarity were filtered out from the dataset

  2. We applied per-video and per-channel clustering and selected the majority speaker cluster only

Please find all the relevant filtering pipeline details in this paper: https://www.isca-archive.org/interspeech_2023/yakovlev23_interspeech.pdf

Best regards,
Anton
antonioperfecto changed discussion status to closed

Sign up or log in to comment