nBest List about vosk-api HOT 11 CLOSED

Sure, but you have to modify recognizer C++ source.

from vosk-api.

Does anyone share the idea of returning nbest outputs?

from vosk-api.

@YunzhaoLu you can check the linked PR to get an idea of how to do it. It might not be the final version though.

from vosk-api.

@sskorol Thank you.

from vosk-api.

We have this now with SetMaxAlternatives method.

from vosk-api.

Hi,

did anyone play with mixing this SetMaxAlternative feature and a next word model predictor to increase accuracy? If yes, any feedback on the soundness of that approach?

E.g. I was transcribing some audio about space exploration in english (although the speaker has a south african accent, guess who that can be...) with vosk api, which is doing a pretty good work. Simply, from time to time for instance it would detect "launderette" instead of "launchpad", and I was thinking that certainly with some context awareness the transcription should be able to prefer "launchpad" as a higher probability than "launderette".

from vosk-api.

Simply, from time to time for instance it would detect "launderette" instead of "launchpad"

You can adjust the language model with https://alphacephei.com/vosk/lm

from vosk-api.

You mean that the model I am using may not contain "launchpad" in its vocabulary and I should add it? I am using the 1.8GB eng model.

from vosk-api.

The word launchpad is already there, it might just have suboptimal probability.

from vosk-api.

Yes so hence why I was looking to use some context awarness model (next word predictor , using BART for example) to pick up lower proba words using a model which tracks the ttranscript domain (space here).
I am not clear what I can do with the language model here though. I understand that the language model tweaking is useful to add vocabulary for abbreviations or proper nouns for example, but in my scenario I am not sure what I can do really. Maybe I am missing something on what LM can bring though.

from vosk-api.

Maybe I am missing something on what LM can bring though.

Yes. LM is exactly about context awareness.

from vosk-api.