Transfer learning

[flassTer]

Is it possible to start transfer learning with more characters in the vocabulary file than the ASR model was already trained on?

[blisc]

We have never tried this. I can only offer some starting points that may or may not work.

• You can try loading the entire encoder weights but create a new decoder from scratch and try training from there
• You can load the entire model and add another layer in the decoder and starting training from there
• You can load the final 1024 -> 29 labels layer and append [1024, num_new_chars] to the weight tensor (I am not sure how you can do this in TF though)
• You can experiment doing any of the above but also freeze the encoder weights, and/or apply a warmup learning rate from 0

Depending on what you are trying to accomplish, it might simply be easier to either hard-code or train a neural LM to convert the model predictions to the desired transcripts. Eg, words -> numbers, dollar -> $, etc.

[flassTer]

Thank you @blisc