search

NVIDIA / OpenSeq2Seq

Toolkit for efficient experimentation with Speech Recognition, Text2Speech and NLP
https://nvidia.github.io/OpenSeq2Seq
Apache License 2.0
1.54k stars 369 forks source link

KenLM intermittent issue #357

Open bmwshop opened 5 years ago

bmwshop commented 5 years ago

On certain samples, intermittently (but more likely on longer ones), when we inference with KenLM, we end up with gibberish at the end of the end of the transcript, eg.

"four one seven crivenuehioforfivzfofofivtresevonfozodichargdignowetnpscosinollservicetdawitworefis...."

this contributes to an overall lower WER score with an LM.

borisgin commented 5 years ago

What was the output of greedy decoder?

SibtainRazaJamali commented 5 years ago

I have the same problem.

SibtainRazaJamali commented 5 years ago

I guess the problem is ctc_decoder.When i made the language model from native client provided by original deepspeech implementation, the results were impressive but the inference speed was very low. But when i built the language model with OpenSeq2Seq toolkit the inference speed is high but wer is also high. I am getting the results like above. e.g on forms that selemtwetysevenfeightenhoe