Markov.py

[Alex Martelli]

"Carel Fellinger" <cfelling at iae.nl> wrote in message
news:94n920$e66$1 at animus.fel.iae.nl...
    [snip]
> > Yep, I remember the debates!-)  The little fact that our
> > approach *WORKED* (allowing a 20,000-word vocabulary real
> > time recognizer to be first built in 1984), of course,
>
> Impressive! (I take it it did single words only, and I guess it needed
> training to get a good profil of the speaker, but still impressive)

Real-time speech but not continuous one (short-as-you-wish,
but mandatory, pause between words -- we called it 'dictation'
speech), and 15-minutes training needed per speaker; took a
dedicated IBM mainframe with a few CPU's and 'vector features',
or equivalent (3rd party vector-processing units for the same
class of mainframe).  It _was_ 1985 before 'we' (actually
some brilliant HW guys in IBM Research) turned out special-
purpose boards that you could stuff inside an IBM PC/AT to
make it perform the same task (maybe even early '86 before
the latter box was reliable enough to announce & demo) --
what we contributed to the miniaturization effort was a
study of numeric sensitivity, to turn the computations into
fixed-point ones at various minimal-sufficient precisions
along the algorithms' paths.


> But if I remember correctly, the use of frequency analysis to
> recognize spoken words was not what was criticised at all. We (the

You don't (remember correctly) -- being the target of such
criticisms sharpens one's memories:-).  I remember silencing
a critic once in public debate about this -- he claimed it's
not truly feasible to 'take dictation' without any understanding
whatsoever, and I gave the counterexample of my father and his
secretary... he, a specialist physician, all the time dictating
extremely specialized and obscure texts -- she, an efficient
dactilographist, not caring AT ALL about understanding of what
she was writing down, as long as she knew exactly how to SPELL
each of the totally-meaningless-to-her words and names.  In fact,
that's why, she explained, she could not do *stenography* for
his reports -- THAT does require a modicum of understanding
of WHAT is being written about, it seems -- rather, she typed
right at the typewriter as he dictated (the way he DID dictate
would have been perfectly suitable for our recognizer, save
for occasional [and inadvertent] lack-of-pause between two
adjacent words... which occasionally made HER go wrong, too!).

> members of the Rosetta team) felt a bit awkward by the claim that
> stochastic models were better suited to attach meaning to (written)
> natural language and even to weed ambigueties.  Grammars, on the other
> hand, are soo crisp and clear, hence easy to understand:) Probably had

Right -- and therefore insufficient to model what human beings
actually DO with language:-).

> to do with some defect in our minds; I for one never grogged Perl, yet
> with Python it was love at first sight.

Same here, though I did use Perl a lot before finally meeting
Python (lack of perceived alternatives...).


Alex