Speed Performance

[pengzhendong]

Excellent work. Have you tested the speed between Reverb and Whisper or Canary-1B?

[jprobichaud]

Thank you! Yes, we did, but it's hard to publish these numbers as so many factors are to be considered here. For example, Whisper has various model sizes available, and different inference implementations are available also.

Also, if you want to perform inference in CPU instead of GPU, the picture changes also.

As a general rule, Reverb is smaller than Whisper medium in terms of number of parameter (and even more so compared to Canary-1B) and thus faster for most application configurations, but again, that can change based on your context.

Do you have a specific comparison you'd be interested in ?

[pengzhendong]

I am interested in the performance as presented in the following paper: https://arxiv.org/pdf/2404.09841.