[ale] sphinx voice recognition

[Bjorn Dittmer-Roche]

On Mon, 20 Oct 2003, Pete Hardie wrote:

> Bjorn Dittmer-Roche wrote:
> > Has anyone used sphinx for normal dictation. As far as I can tell, you can
> > only use it with a very limited set of commands derived from turtle
> > graphics ("Go Left ten", "Rotate one hundred and four degrees" etc) and
> > even then it is not very accurate (although my sound setting smay not be
> > correct). I understand I need a more complete "language model" or
> > "dictionary" or something, but I don't know where I can find that. Perhaps
> > Sphinx in not for such general purpose use? Any help/advice would be
> > great.
>
> You are correct that you need another language model.  On the Sphinx sourceforge
> page there is a perl script called SimpleLM that purports to take a list of
> phrases and create an LM, but the version I downloaded needed a lot of work to
> do it properly (lot of work == rewrite as ksh script/Python module).  That said,
> I was able to build a small corpus and have phrases not in the turtle command
> set be recognized.

Are you able to do dictation with this? Ideally, I would like to use it
for transcription as well, which is harder, and it sounds like it isn't
even up to dictation. I understand sphinx3 is more accurate, do you know
if it uses the same language models?

>
> One thing that I think might still be a stumbling block is the voice itself.  I
> happen to be fairly close to the speaker of the default model, so my utterances
> fall into the same frequencies, etc.  YMMV.
>

Hmm. That could be limiting.

Thanks for your thoughts,

	bjorn