[Tfug] OT? "Alarm/alert" suggestions 8yba0a8e

[Bexley Hall]

Hi Robert,

On 4/11/2014 10:02 AM, Robert Hunter wrote:
> On Thu, Apr 10, 2014 at 9:37 PM, Bexley Hall<bexley401 at yahoo.com>  wrote:
>
>> So, the question(s) I pose is (are):
>> - what sort of "indicator" to listen for?
>> - how to "verify" the intent of that "indicator" (false positives)?
>> - flaws in this approach, in general?
>> - alternative approaches??
>
> This is an interesting problem.

<grin>  Those that aren't are not worth thinking about!  ;-)

> Although I know very little about signal
> processing and speech recognition, I did read recently about Blind Signal
> Separation, and techniques such as Independent and Principle Components
> Analysis.   I wonder if there are implementations that you could adapt for
> your needs.  I would be interested in the results!

Hmmm... this is interesting!  Though the demo appears to have a
flaw -- either in the implementation or the underlying algorithm
(e.g., I selected three different sources and it recovered one
of them cleanly -- while the other two "recovered signals" were
essentially identical mixes of the other two source signals).
I will have to see if I can find more information on their
actual implementation (e.g., how did they position the microphones
and sources? Did they account for time of flight differences or
just intensity?  etc.)

This would require adding at least one more microphone -- presumably
located distant from the first -- *if* the algorithm could magically
lump ALL the "background sound/noise sources" into a single source
(i.e., how would it know that your call for help wasn't part of the
background *mix*?)

But, I'm not even sure what to *detect*!  I.e., a good part of the
problem is figuring out *how* to "signal the need for help" in a way
that you could then algorithmically/mechanically *detect*.

E.g., the classic (advertised) "Help, I've fallen and I can't get up"
tends to rely on the victim pushing a button (warn around their neck,
attached to their wrist, etc.) to initiate the "rescue" series of
events.

Assume, for the moment, the victim is *capable* of pressing that button
(the problem gets markedly more difficult, otherwise).  Who is *likely*
to wear such an "appliance"?  Probably only folks who *expect* the need
for it (e.g., elderly, live-alones, etc.).

So, a solution that relies on wearing that sort of device will only
work -- at most -- for folks who are willing to carry/wear it AT
ALL TIMES.  (I.e., if you just go into the bathroom to wash your
hands you would still need it even if you aren't expecting to
"slip in the shower").

SWMBO has had a couple of occasions in the past where she tried to
call out to me for help but couldn't muster enough "intensity"
to be heard over the noise in my proximity.  On one occasion, she
wasn't even able to call out!

While she or I *might* be willing to carry/wear such a device (but
only while "at home"?), how likely would a healthy teenager?  Age
doesn't innoculate you against the *need* for a rescue mechanism
(a friend -- 30 -- just died in his bathroom while his roommate
was at home).

I thought sound was the most viable approach -- one can *make* sound
a lot easier than being able to actuate a mechanism (unless you will
ALWAYS have that mechanism on your person).  E.g., if you can't talk
(yell), you can always bang on something, etc.

But, there are just too many sounds *in* a bathroom to be able to
easily rule out one vs. another as a valid "indicator" (what if
you are listening to music, "singing in the shower", "body noises",
plumbing noises, etc.)

I think, for example, that the infamous "Clapper" relies on the
fact that percussive events in a home are rare.  And, by requiring
*two* such events in close temporal proximity, the odds of a false
positive are reduced -- while keeping the detection mechanism
simple (I suspect it just looks for a signal that is sufficiently
high above ambient noise level and of short duration... a whistle
might also work to trigger it?)

> Here is a demo of ICA in the Cocktail Party Problem.
>
> http://research.ics.aalto.fi/ica/cocktail/cocktail_en.cgi

Thanks, I'll dig into the math to see what's really going on
under the hood...

--don