TensorSpeech / TensorFlowASR

:zap: TensorFlowASR: Almost State-of-the-art Automatic Speech Recognition in Tensorflow 2. Supported languages that can use characters or subwords
https://huylenguyen.com/asr
Apache License 2.0
931 stars 244 forks source link

Language model with Conformer #196

Open AgaDob opened 3 years ago

AgaDob commented 3 years ago

Hi! Just wondering if the KenLM is compatible with the Conformer model, and if so if there are any example scripts on how to include it? Thank you!

nglehuy commented 3 years ago

@AgaDob Currently KenLM is not compatible with the Conformer or any of Transducer models. LM is not supported for transducer models yet, but I'm working on it and also optimize beam search to use LM.

ashishBT commented 3 years ago

I am curious to know about the integration of language model in your ASR Conformer implementation. As original article of Conformer also mentioned improvement of performance with using LM. You can also assist me, so that I can try.

nglehuy commented 3 years ago

@ashishBT The LM's usually used to rescore the prob in the decoding function of the model (aka beam search). So first we need to optimize the beam search performance and then just connect with LM in the function that calculate the prob (the prob used to get the corresponding class)