Closed zw76859420 closed 1 month ago
Hi Shylock,
For ASR task finetuning, I suggest implementing it using s3prl or espnet. In our paper, we concluded that initializing stage-2 pre-training with stage-1 weights is not beneficial. However, training a stage-1 model is still necessary, as the pre-training target for the stage-2 model is constructed by applying k-means clustering to the features from the stage-1 model.
Thanks for your reply. I understand what you mean.
Dear authors,
Can you tell me how to finetune a model trained in stage-2 for ASR task? We can also draw a conclusion of your paper, for ASR Task, we can directly pretrain stage-2 from scratch without training stage-1?
Looking forward to your reply. Shylock