Just a couple of questions

[mirix]

Hi,

I see you have several models for Speech Emotion Recognition (SER).

Would you say Vesper is the best?

I have also noticed you use acted databases for training, in your experience, does learning between acted and, so called, natural databases transfer well?

I was considering training a (unimodal audio) model on CMU-MOSEI to check if training on a natural database would produce a better performing model in real-life scenarios.

What do you think?

Of course, one could argue that a significant percent of the utterances from YouTube are also acted and do not reflect real emotions, in which case, it it would be better go with professional actors than with amateur Youtubers...

Best,

Ed

[mirix]

I have forked MOSEI to build a unimodal SER dataset:

https://github.com/mirix/messaih/tree/main