KenLM or similar language model

[jesusbft]

Hello,

Do Kraken plan to integrate KenLM or similar language model to help fixing errors in transcriptions? If yes, already have something (codes and ideas) that I could use? PyLaia have something but I never test it.

Thank you,
Weslley O.

[mittagessen]

There is already a beam decoder in `lib/ctc_decoder.py` which is compatible with KenLM but proper language model integration is somewhat tricky. The issue is that decoding happens in label space but the language model works code point space (in whatever granularity it's been trained in). This makes it quite easy to create broken combinations. The new party recognizer on the other hand has much stronger language modeling with its pretrained Llama decoder so there shouldn't be any need for an external LM anymore. It is going to end up in kraken as soon as I figure out how to fine-tune the thing (and got the time to start the integration work).

[jesusbft]

Thank you so much for the informations. I'll take a look in both!!