Datasets:
I want to get the 35 hours labeled text in Greek. How can I set the condition to get the 35 hours labeled Greek text while saving the pseudo labeled text?
I want to get the 35 hours labeled text in Greek. How can I set the condition to get the 35 hours labeled Greek text while saving the pseudo labeled text?
Because when I saved the data and converted it to text, I found that there was a lot of data in the text and some of the audio did not match the voice file.
Hi,
I want to get the 35 hours labeled text in Greek.
This repository contains only the pseudolabels, the entire collection, including the gold labels that you are asking for, is available here: https://github.com/hlt-mt/mosel
Because when I saved the data and converted it to text, I found that there was a lot of data in the text and some of the audio did not match the voice file.
Can you please point out the samples that are affected by this issue?
hello，Thanks for your answer, I have found the problem，There must be some problem with Whisper transcription.For example:
20160118-0900-PLENARY-12_el_1 Υπότιτλοι AUTHORWAVE
20160118-0900-PLENARY-12_el_2 Υπότιτλοι AUTHORWAVE
20160118-0900-PLENARY-12_el_3 Υπότιτλοι AUTHORWAVE
Of course, pseudolabels are not perfect as they are automatically generated and, in the MOSEL case, their quality relies on Whisper quality. This is also discussed in the limitations section of the paper (https://aclanthology.org/2024.emnlp-main.771.pdf).
We double-checked and this is, unfortunately, the case and it is not related to our data generation process. Therefore, I suggest to filter out sentences with this structure (with the "text" field containing "Υπότιτλοι AUTHORWAVE") before using the data.
Thanks for your reply