TCD-TIMIT results about avsr-tf1 HOT 2 CLOSED

Great! Thanks for your contribution.

from avsr-tf1.

Sure, our recent publications include several evaluations on the TCD-TIMIT dataset, where we also mention what architectures we used:

https://georgesterpu.github.io/publications.html

We have an audio-only and audio-visual evaluation on the speaker-dependent set at the character/word level in our ICMI 2018 article.
The ICIP 2018 one has a visual-only evaluation at the viseme level, yet the experiments were performed last year with slightly older architectures. With the ResNet CNN, viseme accuracy should be around 71%.

Although unpublished, when decoding characters directly from the visual stream alone should give a ~45% character error rate, ~72% word error rate.

It is a good idea to create a benchmark page.

from avsr-tf1.