I come again~
Now I’m continuing train with the pre-trained model.
I have audios about 22 hours, and the train parameters are:
Learning-rate: 0.0001
Dropout-rate:0.35
Language model is combining my own train transcript and librispeech’s corpus.
After 75 epochs, the validation loss is 37, train loss is 35, but the test result is not well, though the WER is 0.19, CER is 0.1, test loss is 22.
Now, what should I do？ Can I decrease the loss into a number less than 10？
My audio file is not clean, parts of them have noise.
You need to study how loss on training / validation evolves over time, not their value
Generally speaking people doing transfer-learning have had to use lower learning rate. You need to experiment yourself, there’s no one-value-fits-all we can recommend.
Thank you for reply~