base model encoder choice
#8
by sugintama - opened
For base model's encoder choice, is an encoder with a different FastConformer-like structure compatible with this SALM, such as nvidia/parakeet-tdt-0.6b-v2, or must it strictly be an encoder which is combined with a transformer decoder in base model?
For Canary-Qwen-2.5B specifically, it has to be the encoder architecture and parameters in this checkpoint, otherwise it won’t work.
But if you wanted to train your own SALM with NeMo, you can or course use any pretrained model (just change „pretrained_asr” to a different name or .nemo checkpoint)