1
$\begingroup$

The Figure 2 of the Whisper paper is comparing WER on LibriSpeech clean-dev vs Average WER on other datasets. I believe clean-dev is the typo of clean-test dataset as specified in the Table 9 and Figure 2 is the plots of the Table.

I suppose Figure 2 is demonstrating robustness on out-of-distribution datasets on y-axis vs that on in-distribution datasets on x-axis (LibriSpeech test-clean).

The question is why the other datasets on y-axis listed in the Appendix A can be said out-of-distribution? Are the well known as such? And why the in-distribution vs out-of-distribution is ideal if it is one to one linear relation? Out of distribution results would be difficult to get them right, hence the linear line can be one to N?

enter image description here

enter image description here

enter image description here

enter image description here

enter image description here

$\endgroup$

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.