Train Conformer-CTC ASR model on character tokenzier

[neso613]

Following this doc to fine-tune conformer-ctc model.

In this, subword tokenzier used for conformer-ctc and character tokenzier used for quartznet.

can we use character tokenzier for conformer-ctc model?
If yes, how?

[titu1994]

You can try, but it will probably not be very stable so you'll need to update hyper params. We don't experiment with char models unless they are for Asian languages with large vocabularies

[bene-ges]

@neso613, @titu1994 I recently tried this nemo config for training char conformer-ctc from scratch. It works ok as it is, fast, stable and good quality.

[titu1994]

That's good news ! Any changes to LR or weight decay ?

[bene-ges]

no changes, just default config
the simplest test for English was on an4 dataset,
for Russian language I used a larger dataset, but just replaced the alphabet