On Wed, 10 Sep 1997 14:26:21 -0400 Pierce Nichols <pn30@COLUMBIA.EDU>
writes:

>        DTMF decoders are cheaper than most PICs, I think, but I am
>not
>sure. Doing it with a PIC is awfully klugey, but it might be a fun
>thing
>to try...
>

I've tried the method outlined by Scott Dattalo on his web page.  It does
work, but it requires nearly all the processing power of a 20 MHz PIC in
order to detect any of the 8 tones arbitrarily.  Since I was hoping to
use a 16F84, I abandoned the project.  I could push a 16F84 to 16 MHz but
it definitely wouldn't do 20.  My implementation also needed about 50
bytes of RAM.

It is feasible and may be good for some applications.  Overall I'd
recommend using a decoder chip, especially if the tone input is not going
to be perfectly clean.  Though I only tested the PIC decoder with a
direct connection to a tone telephone, I wouldn't expect it to work very
well with noisy tones.

The method is as follows.  All of the processing can be done in a timer
interrupt at about 15 KHz, so the tone decoder can operate independent of
the main program other than taking a lot of its CPU cycles away.  Each of
the processing steps below is serviced once per interrupt.

Create a local tone at each of the 8 frequencies (4 rows and 4 columns).
I used DDS with 16-bit phase accumulators.  Every interrupt, add a
constant (tone frq * 65536 / interrupt frq) to the accumulator.  The MSB
of the accumulator will be a square wave at the tone frequency.  The MSB
xor the second most-significant bit will be a square wave at the tone
frequency, but quadrature in phase to the MSB (the usefullness of this is
in Scott's description)

Sample the input (1 bit).  I used 3 gates of a 4069 with the first one
biased for linear operation and capacitively coupled to the phone.  This
processes the tone into a square wave suitable for a PIC input.

Exclusive-or the input sample with all the tones, and the quadrature of
all the tones.  If the input contains a signal near the tone frequency,
some of these exclusive-or results will contain a predominance of ones
(or a predominance of zeros if out of phase)

Filter the exclusive-or results.  I used a simple filter which was an 8
bit variable which increased by 1 if the xor was a match, and decreased
by 2 if it didn't.  Half of the filters consider a result of 1 to be a
match, the other half use the same results but call 0 a match.  There are
a total of 32 filters, 4 for each tone.

The main program looks at the filter results, and if any one of the 4
filters for a particular tone is higher than a threshold (I used 32 IIRC)
that tone is present in the input signal.  The filtering needs to be
improved.  Sometimes the tone would drop out periodically.  Lowering the
threshold some to prevent that caused other tones to be falsely detected.

If your application doesn't need to detect all 8 tones simultaneously
then it can be optimized to use a lot less PIC resources by only looking
for one tone at a time.  An example of this would be an answering machine
which listens for it's owner to press 1-2-3-4 to replay any messages
remotely.  The program would look only for the two tones that represent
button #1 at first, if those are detected it would retune for tone 2,
etc.  If no tone in the sequence is detected within a timeout time it
would go back to listening for the first one.  This could be done easily
by running 2 tone generators and 8 filters and changing the DDS constants
depending on which tone is expected.