Re: spandsp tone_detect.c goertzel perodogram

On Oct 25, 10:35 am, Steve Underwood <ste...@xxxxxxx> wrote:
Kamanashis wrote:
Hi,

Let me explain the situation, I am trying to differ beep from
different answering machine.
The information about the answering machine beep frequency is given
below.

[T-Mobile]
frequency=1250 ; frequency in Hz
energy=-30 ; energy in db
duration=250 ; duration in milisecond

[Verizon]
frequency=1500 ; frequency in Hz
energy=-20 ; energy in db
duration=550 ; duration in milisecond

[Sprint/Nextel]
frequency=1400 ; frequency in Hz
energy=-30 ; energy in db
;duration=600 ; duration in milisecond
duration=120 ; duration in milisecond

[AT&T/Cingular]
frequency=1700 ; frequency in Hz
energy=-20 ; energy in db
duration=300 ; duration in milisecond

I tried with goertzel algorithm (I used the implementation from
spandsp library).
It gave me farely close values of energy [T-Mobile], [Verizon] and
[Sprint/Nextel].
It was the [Sprint/Nextel] that I was testing. Then I tested Verizon.
I did not find a
way to detect the amds.

Now I have a few questions,

1. what does they mean by -20db and what is its relation with
goertzel energy.

It means they send the tone at -20dB, but the energy at the receiver
will depend on the losses between the transmitter and receiver.

2. how can I use the periodogram functions of spandsp library, is
there any chance
that it may work for me ?

It could, if you add some more bits and pieces to what is there. The
periodogram code in spandsp is a work in progress for a full supervisory
tone detection scheme.

3. Is there any relation of duration of samples and goertzel
algorithm. (I get frames
of 20ms each, with 160 samples).

The Goertzel algorithm just evaluates a single output bin of a discrete
Fourier Transform. No more, and no less. It, therefore, behaves just
like any other kind of Fourier transform. Basically, the narrower you
want the frequency window of the detector to be, the more samples you
need to put through the Goertzel algorithm. If you are trying to use the
detector for a single tone, and reject voice, you will probably need to
apply windowing to the data, and use a narrow detection range. Don't
make the range too narrow, through, as there will be some tolerance on
the tone's frequency.

And please suggest me what to do to succeed this project.

I don't understand what you are trying to achieve. Are you trying to
build a generalised answering machine detector? Are you trying to
discriminate which of the above services you have connected to?

Detecting telephony tones is trivial. The tricky part is avoiding false
detections of voice as tone. Discriminating which of the services you
have attached to should be easy, as their tones appear distinctive.
Trying to reliably determine if you have connected to an answering
service is much harder.

Regards,
Steve

Thank you for your reply.

I am currently in a project to modify the asterisk amd application
that
detects answering machine depending on the silence period.

I have noted that amd machines say something like "If you want to
leave
a message, please say your message after the beep" and then they
produce a beep. The silence before the beep is greater than 700ms
most of the times which is greater than silence between words. Now
that I can predict that it is beep, I get the goertzel energy of the
given
machines and try to identify a specific machine. As the specifications
are read from a configuration file(not hard coded) so, it should work
for other answering machines too.

Now the specification, like frequency and power, is not verified. And
I
have tried a lot with this specification. Let me share you a little.
I was testing the following,

[Sprint/Nextel]
frequency=1400 ; frequency in Hz
energy=-30 ; energy in db
duration=600 ; duration in milisecond

And I got a statistics as follows(it tells about 20ms, other frames of
20ms gives almost the same result)

tone :goertzel energy

50 :98675440.000000
100 :36266448.000000
150 :14138939.000000
200 :277978.187500
250 :20917570.000000
300 :93951144.000000
350 :221786384.000000
400 :306548128.000000
450 :677073088.000000
500 :1453333888.000000
550 :2423981056.000000
600 :5636514304.000000
650 :9835254784.000000
700 :58524889088.000000
750 :247646289920.000000
800 :5774915072.000000
850 :1130472576.000000
900 :118388688.000000
950 :26353552.000000
1000 :28255540.000000
1050 :127593568.000000
1100 :149558512.000000
1150 :216797728.000000
1200 :217243552.000000
1250 :289937792.000000
1300 :318662848.000000
1350 :297567552.000000
1400 :296347712.000000
1450 :266078320.000000
1500 :230130080.000000
1550 :167810704.000000
1600 :165656928.000000
1650 :95268592.000000
1700 :52679632.000000

Now, the question is, is there any mistake appearent ?
I mean it seems that the peak is in 750 frequency all the
time whereas I am expecting it to be in 1400. Is there any
possible mistake in my calculation ? I am confused with
the data I get.

I tried with beep of,

[Verizon]
frequency=1500 ; frequency in Hz
energy=-20 ; energy in db
duration=550 ; duration in milisecond

It gave me almost the same result(750Hz got the highest
energy).

-- Thanks
Kamanashis Roy

.

Relevant Pages