ULCA-asr-dataset-corpus icon indicating copy to clipboard operation
ULCA-asr-dataset-corpus copied to clipboard

Published 20 hours ago • verified publisher icon Open-Speech-EkStep

Vakyansh models are fine-tuned on much less/more labelled data (transcribed audios). So why there is a discrepancy?

Open ritwikmishra opened this issue 3 years ago • 0 comments

In this repo the size of Hindi labelled data is mentioned as 4200 hours but this repository contains Hindi labelled data with duration of 2398 hours.

On the contrary, Tamil labelled data in this repository (1160 hours) is much more than Tamil data (250 hour) on which Vakyansh model for Tamil has been fine-tuned.

Similarly for other languages

  • Kannada here (608 hours) vs Vakyansh Kannada (560 hours)
  • Bengali here (541 hours) vs Vakyansh Bengali (200 hours)
  • etc

So why there is discrepancy?

ritwikmishra avatar Feb 15 '22 08:02 ritwikmishra