0

How can I integrate a language model in a tensorflow speech recognition architecture?

There are a bunch of examples out there for building character level speech recognition in Tensorflow (e.g. https://github.com/nervanasystems/neon, https://github.com/buriburisuri/speech-to-text-wavenet), which is interesting but practically useless, unless a language model is integrated. I couldn't find an example that uses a language model.

How can I integrate a language model?

Nikolay Shmyrev
  • 24,897
  • 5
  • 43
  • 87
bear
  • 663
  • 1
  • 14
  • 33

1 Answers1

2

LM scoring is just an additional rescoring step, simply a spelling correction with a language model. It can be applied on any system output. Mozilla has it spell.py for example.

Nikolay Shmyrev
  • 24,897
  • 5
  • 43
  • 87
  • 1
    The file was removed with newer commits and in some branches. the file can be found in this [branch](https://github.com/mozilla/DeepSpeech/blob/deep_compression/util/spell.py) – JorgeC Dec 04 '17 at 17:42