finetune the decoder on text only data?
#7
by
leestevennz
- opened
Hi there,
Just wondering if it possible to finetune the decoder on text only data, for domain adaptation?
Canary's decoder is a Transformer LM conditioned on encoder outputs so it is possible to adapt the decoder of a sequence-to-sequence ASR model like Canary using only text data.
It can be done through shallow or cold/deep fusion, or (what I would recommend) continued pretraining.