Fine Tuning Language Model

[sruteeshkumar]

Hi,
I would like to fine tune the language model to decode my domain specific vocab.
I referred to the Readme at https://github.com/mozilla/DeepSpeech/blob/master/data/lm/README.md

I have domain specific text of about 1 Lakh sentences which is much smaller compared to the Common Voice Text and the Libri Texts. This would not modify the langauge model much.

I want my ASR model to be able to decode generic english as well as my domain specific vocab.

Is there any way I can use different weights to the generic english texts and to my domain specific texts while training the language model.

Thanks.

[lissyx ((slow to reply) [NOT PROVIDING SUPPORT])]

@sruteeshkumar

Is there any way I can use different weights to the generic english texts and to my domain specific texts while training the language model.

Something like Allow use of several decoders (language models) with a single model in the API · Issue #1678 · mozilla/DeepSpeech · GitHub ?

[sruteeshkumar]

Hi
Yeah something similar. Is there any reference I can look into for building this?
Ideally I would want a single model trained using both the texts.
Thanks.

[lissyx ((slow to reply) [NOT PROVIDING SUPPORT])]

Check the issue, we don’t yet support that.

[sruteeshkumar]

@lissyx Yes, I saw the issue. Do you have any references which I can look into so that I can build it myself.
Thanks.

[lissyx ((slow to reply) [NOT PROVIDING SUPPORT])]

@sruteeshkumar

Do you have any references which I can look into so that I can build it myself.

I don’t understand your question: what do you want to build yourself?

[sruteeshkumar]

Build the language model with different weights for both the texts (generic english and domain specific texts). Any references/ pointers for this is appreciated.