Number conversion: "seven twenty two" becomes "7 twenty-two"

[KJ7LNW]

When I say "seven twenty two" it becomes "7 twenty-two", but I would expect "722".

I also tried:

• "seven hundred twenty two" becomes "700 twenty-two"
• "seven hundred and twenty two" becomes "700 and twenty-two"

[ideasman42]

I can't redo this:

• "seven hundred and twenty two" becomes "722" here.

Other combinations noted are more ambiguous, what is your --numbers-min-value set to?

[KJ7LNW]

Oops, funny, I wrote the min value stuff and forgot it was there. I have --numbers-min-value=3 which explains the issue.

If you see an easy way to make --numbers-min-value realize that 722 is much bigger than 3, then go for it because I'm not quite sure how to hook that in properly. Otherwise, you may close the issue as not a bug.

[ideasman42]

Could you show the command used to activate the nerd-dictation ? I can't redo the issue even with --numbers-min-value=3 set.

[KJ7LNW]

 ./nerd-dictation begin --numbers-as-digits --numbers-no-suffix --numbers-min-value=3 --suspend-on-start --verbose=1 --simulate-input-tool=DOTOOLC

[ideasman42]

I still can't redo this even with the exact command. It might be the language model outputs "twenty-two" instead of "twenty two" which confuses the parsing - which assumes spaces.

The readme in the model directory reads:

Accurate universal English model (both for callcenter and wideband)

Based on Appen Kaldi model https://github.com/Appen/UHV-OTS-Speech

Librispeech test-clean WER:  5.69%
Tedlium WER:                 6.05%
Callcenter WER:             29.78%

[KJ7LNW]

That could be. I'm using an unmodified vosk-model-en-us-0.42-gigaspeech model and it does say twenty-two.