Initialize a Parakeet Cache-Aware Streaming model's encoder from an offline model.

[gabitza-tech]

Hello,

Thank you for all your amazing work!

At the moment, nemo only support the fastconformer 120M streaming model, however I would like to also try a bigger model. I was wondering if it is possible to modify the architecture to match de 600M parakeet model and initialize the encoder weights from the offline parakeet model? Otherwise, training a bare 600M streaming model without any SSL pretraining would probably be worse and slower to train than the 120M model.

Best regards,
Gabi