Interesting problem can you add more detail on these two points

1. Multiple conversations

More than one operator on a single line? How does this work?

2. Recovering letters from context

This is the ultimate goal? Are letters omitted in a consistent formalized
way (e.g. can't) or in an informal, individual way (i.e. wut, wht? w?)




On Mon, Jul 21, 2014 at 11:35 AM, Chris Albertson <[email protected]
> wrote:

>
>
> I'm completely new to nupic.  However I have a few decades experience in
> software AI and related fields.  What I'm asking is if Nupic is something I
> should look into.  I've read "On Intelligence" and some of the web page but
> the Nupic software seems to be a tiny subset of that.
>
> The subject area is Morse Code recognition.  In the simple case of machine
> generated Morse Code with no noise this is a very easy problem.  But in the
> real world the timing is variable,there is VERY high levels of noise and
> likely several conversations all going on at once.
>
> My task is about like going into a noisy bar where many people are talking
> loud and attempting to simultaneously transcribe every conversation.  Ok,
> it's easier because Morse is almost already digital and the data rate is
> only as fast as human hand can push a single contact key.
>
> I can reduce the date rate by time tagging the transitions when a tone
> changes.  This reduces the data from the 48,000 sample per second audio to
> about 100 or so tokens per second.  But I would MUCH prefer to use a
> learning algorithm to pick these transition out of the noise.  Best if I
> could input the audio directly. (Yes I saw the audio example)
>
> The problem is very much like speech.  It is hierarchical with "elements"
> (dots , dashes, spaces,..) making up letters and letters making words and
> words making up very stylized phrases.   Also like speech humans are
> "sloppy" and slur it and change their speed and talk over top of one
> another.
>
> My plan currently to using well known speech recognition algorithms but
> hand tuned to Morse Code.
>
> What I'm wondering after reading On Intelligence is if NuPic is up to the
> task. I know I'd have to build in my own hierarchy but the levels are so
> well defined.    The goal is to recover the letters that are sent using
> context to "fill in the gaps"  but NuPic does not seem like something you
> train and I don't see how it could ever recover the letters and words.
> Am I correct to assume CLA might work but NuPic is not up to this task?
>  I'd be happy if I were wrong.
>
> --
>
> Chris Albertson
> Redondo Beach, California
>
> _______________________________________________
> nupic mailing list
> [email protected]
> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
>
>
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to