[clug] linux speech to text? [SEC=PERSONAL]

Daniel Pittman daniel at rimspace.net
Sun Sep 20 20:31:42 MDT 2009


"Roppola, Antti - BRS" <Antti.Roppola at daff.gov.au> writes:

> I recall a couple of years back IBM open-sourced their speech recognition
> software: http://linux.slashdot.org/article.pl?sid=04/09/28/2226253
>
> I've known a few people who have had RSI issues and nearly all have ended up
> using the proprietary Dragon Naturally Speaking product.

*nod*  As I understand it, this is still the gold-grade of speech-to-text.

> This software requires trahing to recognise an individual's idiosyncracies,
> high grade audio hardware, measured speech patterns & low ambient
> noise. Even then the results are sometimes amusing.

I ask, because my partner looked into it some years back due to RSI issues and
it was pretty dreadful at the time, but claims are that it has progressed to
speaker-independent recognition without training these days.  When did y'all
last test out the product?

> I wonder just how well alternatie products might handle audio without these
> pre-conditions (i.e. random people captured on a cheap mic talking at 100 to
> the dozen in a noisy room).

I understand that this is more or less where things are getting to.


> I was watching TV at a café last week and they had the closed captions on as
> it was too noisy to hear. The captions are obviously automatically generated
> as some of the subtitles were rather amusing. On the whole, the results were
> pretty impressive considering the above factors.

Nope: those captions are human-generated, and you can sometimes see the live
edits as they correct it.  Typically, though, that is someone typing to keep
up with speech, using specialized keyboards and input equipment, and often
with less than seven seconds delay between capture by the camera and it
landing on your screen.

The error rates are surprisingly low given the constraints they are operating
under, actually. :)

Regards,
        Daniel

-- 
✣ Daniel Pittman            ✉ daniel at rimspace.net            ☎ +61 401 155 707
               ♽ made with 100 percent post-consumer electrons
   Looking for work?  Love Perl?  In Melbourne, Australia?  We are hiring.


More information about the linux mailing list