Automatic Gain Control

Hi All,

  I wish to make a AGC system targeted for decoded speech signals in
mobile phones. I went through many of Jerry Avin's previous posts on
this topic and tried to work in those directions. However things are
not that easy when it comes to real time controlling of gain of the
speech singals.
From the previous posts I can summarise following:

 1) If the gain of the singal is higher than the average level at
which we want to operate, we need to decrese the gain. This decrease
has to happen quickly.
 2) If the signal gain is lower than some threshold, we need to
increase the gain. This has to be done gradually.
 3) There are some dead zones where no gain control be applied.


 Now my doubts are as follows:

1) Which method should be used to analyse the signal level. Is peak
detection the right method of estimating the signal energy variation,
or the average / moving average (with a history of say prev 10 frames/
1600 samples)of the singal energy gives the right picture.

2) What are dead zones? Do they mean the pauses in the speech signal
where the speaker does not speak anything? Should AGC be active during
pauses?
As far as I conceive, AGC should remain active during pauses because
that would ensure no discontinuity between the speech segment and the
pauses.
If we keep AGC ON guring some speech segment and make it OFF as soon
as we detect that we have silence, that would create some sort of
discontinuity in the overall volume of the flow of the speech. So the
noise during the pauses hsould also be corrected accordingly.

3) Shouldn't our decision making mechanism take into account value of
noise floor.

 Could some one please explain to me how do I go about this. I am
finding this all too tricky. Each time I have to tune my attack rate
or the threshold or the noise floor level. It works sometimes for some
speech signal BUT then sounds arbit for some other speech signal. I am
a bit confused and dont know hoe to go about it. I need your help with
this.

Just for your information I have set following values for my
experiment:
* a FRAME is 20ms long consisting of 160 samples.

a ) ATTACK_SLOPE	   -0.0322 dB per frame (frame = 160)samples
b ) HANG_OVER_PERIOD      30 frames (Observation time where u analyze
and make some decision).
c ) NOISE_LEVLE	         -95 dB. (avg. noise floor level)
d ) TOLERANCE_LEVEL	 20 dB (The diff between noise floor and peak
detection curve. Anything less than this should be treated as noise.)


Thanks and Regards,
Anoop Deoras
Sasken, India
Email : adeoras [at] gmail [dot] com

"EC-AKD" <akd_ecc@yahoo.com> wrote in message
news:2f592dd3.0503252254.79c9e214@posting.google.com...
> Hi All,
>
>   I wish to make a AGC system targeted for decoded speech signals in
> mobile phones. I went through many of Jerry Avin's previous posts on
> this topic and tried to work in those directions. However things are
> not that easy when it comes to real time controlling of gain of the
> speech singals.
> From the previous posts I can summarise following:
>
>  1) If the gain of the singal is higher than the average level at
> which we want to operate, we need to decrese the gain. This decrease
> has to happen quickly.
>  2) If the signal gain is lower than some threshold, we need to
> increase the gain. This has to be done gradually.
>  3) There are some dead zones where no gain control be applied.
>
>
>  Now my doubts are as follows:
>
> 1) Which method should be used to analyse the signal level. Is peak
> detection the right method of estimating the signal energy variation,
> or the average / moving average (with a history of say prev 10 frames/
> 1600 samples)of the singal energy gives the right picture.
>
> 2) What are dead zones? Do they mean the pauses in the speech signal
> where the speaker does not speak anything? Should AGC be active during
> pauses?
> As far as I conceive, AGC should remain active during pauses because
> that would ensure no discontinuity between the speech segment and the
> pauses.
> If we keep AGC ON guring some speech segment and make it OFF as soon
> as we detect that we have silence, that would create some sort of
> discontinuity in the overall volume of the flow of the speech. So the
> noise during the pauses hsould also be corrected accordingly.
>
> 3) Shouldn't our decision making mechanism take into account value of
> noise floor.
>
>  Could some one please explain to me how do I go about this. I am
> finding this all too tricky. Each time I have to tune my attack rate
> or the threshold or the noise floor level. It works sometimes for some
> speech signal BUT then sounds arbit for some other speech signal. I am
> a bit confused and dont know hoe to go about it. I need your help with
> this.
>
> Just for your information I have set following values for my
> experiment:
> * a FRAME is 20ms long consisting of 160 samples.
>
> a ) ATTACK_SLOPE    -0.0322 dB per frame (frame = 160)samples
> b ) HANG_OVER_PERIOD      30 frames (Observation time where u analyze
> and make some decision).
> c ) NOISE_LEVLE          -95 dB. (avg. noise floor level)
> d ) TOLERANCE_LEVEL 20 dB (The diff between noise floor and peak
> detection curve. Anything less than this should be treated as noise.)
>
>
> Thanks and Regards,
> Anoop Deoras
> Sasken, India
> Email : adeoras [at] gmail [dot] com

Assume yin is the input to the AGC (say a speech signal) and yout is the AGC
output.

loop forever
{
/* Voltage controlled amplifier is just a multiplier here */
yout=yin*iout

/* error */
err=spoint-abs(yout)

/* Integrate */
iout1=iout
iout=iout1+gain*err }




Where in the above 'gain' needs to be found by experiment. Too large
and the envelope will be compressed and the loop will go unstable. If the
gain
is too low then the loop will be slow to repond. Try gain=0.001 for starters
though
the gain value for stability depends on the size of yin. spoint is a set
point and
can be taken to be 1.0.   yout is the AGCd output.

regardsDr Tam

in article 2f592dd3.0503252254.79c9e214@posting.google.com, EC-AKD at
akd_ecc@yahoo.com wrote on 03/26/2005 01:54:

> I wish to make a AGC system targeted for decoded speech signals in
> mobile phones. I went through many of Jerry Avin's previous posts on
> this topic and tried to work in those directions. However things are
> not that easy when it comes to real time controlling of gain of the
> speech singals.
> From the previous posts I can summarise following:
> 
> 1) If the gain of the singal is higher than the average level at
> which we want to operate, we need to decrese the gain. This decrease
> has to happen quickly.

this means a fast attack time constant ...

> 2) If the signal gain is lower than some threshold, we need to
> increase the gain. This has to be done gradually.

... and a slow release time constant

> 3) There are some dead zones where no gain control be applied.

that gets defined in your compression curve.


> Now my doubts are as follows:
> 
> 1) Which method should be used to analyse the signal level. Is peak
> detection the right method of estimating the signal energy variation,
> or the average / moving average (with a history of say prev 10 frames/
> 1600 samples)of the singal energy gives the right picture.

it depends on what you're trying to accomplish with your AGC.  if the
purpose was to adjust the level so that clipping does not happen, then peak
detection might be appropriate.  if the purpose is to even out the loud and
quiet segments of sound so that a person can hear it better, then some dB
loudness detection might be the right thing.  that might be a mean square
measure (energy) and perhaps with a perceptual weighting filter (maybe "A
weighting") applied before that.   you can also make it some combination of
the two, peak level and perceptual loudness detection.


> 2) What are dead zones? Do they mean the pauses in the speech signal
> where the speaker does not speak anything? Should AGC be active during
> pauses?

besides limiting at the top end, you might want to have a gate at the bottom
end so that ambient noise doesn't get transmitted when the speaker isn't
talking.  maybe not.  i dunno what your specs or application is.  (radio
people used to call this "squelch".  audio folks call it "gating".)

> As far as I conceive, AGC should remain active during pauses because
> that would ensure no discontinuity between the speech segment and the
> pauses.

you can have it kick in and out smoothly.  in fact, strictly speaking, it is
always active, even when the gain is unaffected, as long as it is measuring
level and *will* react when your level thresholds are crossed.

> If we keep AGC ON guring some speech segment and make it OFF as soon
> as we detect that we have silence, that would create some sort of
> discontinuity in the overall volume of the flow of the speech.

it can (and should) be done smoothly.  that is part of the calculus of
attack and release time.

> So the
> noise during the pauses hsould also be corrected accordingly.
> 
> 3) Shouldn't our decision making mechanism take into account value of
> noise floor.

if you have the means of measuring it.  perhaps you want the user to specify
it.

> Could some one please explain to me how do I go about this. I am
> finding this all too tricky. Each time I have to tune my attack rate
> or the threshold or the noise floor level. It works sometimes for some
> speech signal BUT then sounds arbit for some other speech signal. I am
> a bit confused and dont know hoe to go about it. I need your help with
> this.
> 
> Just for your information I have set following values for my
> experiment:
> * a FRAME is 20ms long consisting of 160 samples.
> 
> a ) ATTACK_SLOPE       -0.0322 dB per frame (frame = 160)samples

i think you should consider these parameters in units that do not depend on
your sample rate.  something per second.  then, knowing the sample rate, map
that to something per sample for your implementation.

> b ) HANG_OVER_PERIOD      30 frames (Observation time where u analyze
> and make some decision).
> c ) NOISE_LEVLE             -95 dB. (avg. noise floor level)
> d ) TOLERANCE_LEVEL     20 dB (The diff between noise floor and peak
> detection curve. Anything less than this should be treated as noise.)

so how do you want to treat noise?  gate it to zero?  or maybe to some other
reduced gain?  or leave it alone?


even if you process "frames" or "segments" or "blocks" or "chunks" or
"buffers" (there are lotsa names for it) of audio, your algorithm should be
a samply-by-sample algorithm to make sure its behavior is smooth.  other
than a throughput delay, it shouldn't matter to the sound what blocksize you
end up using (increased blocksize *might* decrease computational expense at
the cost of increased delay).

are you doing this in MATLAB or C or something like that for simulation?


-- 

r b-j                  rbj@audioimagination.com

"Imagination is more important than knowledge."

you want to do some googling using terms like "audio compressor"
"dynamic range compression" .  See some of the products made by dBx
such as the dbx160.  also look up  "orban optimod"

heres more, some may be old but look around there ...
http://www.harmony-central.com/Effects/effects-explained.html
http://www.harmony-central.com/Effects/Articles/Expansion/
http://www.harmony-central.com/Effects/Articles/Compression/
http://www.alesis.com/  (see beginners guide to compression)


Mark

"Mark" <makolber@yahoo.com> wrote in message 
news:1111897578.535028.149910@f14g2000cwb.googlegroups.com...
> you want to do some googling using terms like "audio compressor"
> "dynamic range compression" .  See some of the products made by dBx
> such as the dbx160.  also look up  "orban optimod"
>
> heres more, some may be old but look around there ...
> http://www.harmony-central.com/Effects/effects-explained.html
> http://www.harmony-central.com/Effects/Articles/Expansion/
> http://www.harmony-central.com/Effects/Articles/Compression/
> http://www.alesis.com/  (see beginners guide to compression)

I'll add one more:
http://www.rane.com/pdf/note141.pdf 

"EC-AKD" <akd_ecc@yahoo.com> wrote in message 
news:2f592dd3.0503252254.79c9e214@posting.google.com...
> Hi All,
>
>
> 2) What are dead zones? Do they mean the pauses in the speech signal
> where the speaker does not speak anything? Should AGC be active during
> pauses?
> As far as I conceive, AGC should remain active during pauses because
> that would ensure no discontinuity between the speech segment and the
> pauses.
> If we keep AGC ON guring some speech segment and make it OFF as soon
> as we detect that we have silence, that would create some sort of
> discontinuity in the overall volume of the flow of the speech. So the
> noise during the pauses hsould also be corrected accordingly.

Sometimes you want to freeze the changes in the AGC, keeping the gain constant. 
This is sometimes called "hold".  This can be used so as not to crank up the 
volume unduly during "silence" (which in reality, will probably be the noise 
floor and not digital zeros). 

One thing to be aware of is "comfort noise".  If you detect
low noise=non-speech don't set it to zeros but add a little
white noise to let the listener know that the line is still
active and not disconnected.  Morphing from this "comfort
noise" level back to speech and vice-versa is much less
jarring than to zero level.  Sometimes "Noise" is a good
thing.

-- 
Chip Wood

"Jon Harris" <goldentully@hotmail.com> wrote in message

> > 2) What are dead zones? Do they mean the pauses in the
speech signal
> > where the speaker does not speak anything? Should AGC be
active during
> > pauses?
> > As far as I conceive, AGC should remain active during
pauses because
> > that would ensure no discontinuity between the speech
segment and the
> > pauses.
> > If we keep AGC ON guring some speech segment and make it
OFF as soon
> > as we detect that we have silence, that would create
some sort of
> > discontinuity in the overall volume of the flow of the
speech. So the
> > noise during the pauses hsould also be corrected
accordingly.
>
> Sometimes you want to freeze the changes in the AGC,
keeping the gain constant.
> This is sometimes called "hold".  This can be used so as
not to crank up the
> volume unduly during "silence" (which in reality, will
probably be the noise
> floor and not digital zeros).
>
>

robert bristow-johnson <rbj@audioimagination.com> wrote in message news:<BE6B7E38.59CF%rbj@audioimagination.com>...
> in article 2f592dd3.0503252254.79c9e214@posting.google.com, EC-AKD at
> akd_ecc@yahoo.com wrote on 03/26/2005 01:54:
> 
> > Now my doubts are as follows:
> > 
> > 1) Which method should be used to analyse the signal level. Is peak
> > detection the right method of estimating the signal energy variation,
> > or the average / moving average (with a history of say prev 10 frames/
> > 1600 samples)of the singal energy gives the right picture.
> 
> it depends on what you're trying to accomplish with your AGC.  if the
> purpose was to adjust the level so that clipping does not happen, then peak
> detection might be appropriate.  if the purpose is to even out the loud and
> quiet segments of sound so that a person can hear it better, then some dB
> loudness detection might be the right thing.  that might be a mean square
> measure (energy) and perhaps with a perceptual weighting filter (maybe "A
> weighting") applied before that.   you can also make it some combination of
> the two, peak level and perceptual loudness detection.
> 
I calculated the Peak level and also the moving average of energy of
the signal (with a history of past 1.5 sec speech segment). Taking
average of these two I can now crudely see the general behaviour of
the speech. However still deciding on some threshold value is tough.
Sometimes I feel that I may also need to tune the threshold values
depending upon the speech I am working on.

I am coding in C and use matlab for plotting the curves.

I generated a speech waveform by talking into a microphone from some
distance and recorded it.. after some time I gradually came closer to
the microphone so as to increase the overall energy. Now my aim is to
give a boost to that signal which was recorded from some distance.
Such would be a scenario when the user would use a hands free mobile.
I am assuming that the change from faded speech to loud speech is not
drastic and not much variaton is there.. as in a speech seg of 1 min,
30 sec would be faded speech and remaining 30 sec would be loud
speech.
Initially to get going, I wish to correct the faded 30 sec speech by
assuming some thresholds (some dB value). Later I may be able to tune
the thresholds properly.

As I am doing it all alone, I feel lost when I dont see it going
anywhere. Suggestions and help from usenet would be really
appreciated.

Thanks 
Anoop

in article 2f592dd3.0503300144.7544fb03@posting.google.com, EC-AKD at
akd_ecc@yahoo.com wrote on 03/30/2005 04:44:

> As I am doing it all alone, I feel lost when I dont see it going
> anywhere. Suggestions and help from usenet would be really
> appreciated.

can't help much at a distance.  i can only suggest, if you have the sound
recorded on some kind of wave editor (like Cool Edit), that you select the
quiet pieces and manually boost them to what sounds right, then look at your
measured loudness and the dB boost for that section and see if you can
construct a heuristic curve that maps dB loudness (relative to fullscale) to
dB of boost.

rots o' ruk.

-- 

r b-j                  rbj@audioimagination.com

"Imagination is more important than knowledge."

And of course, most editors like Cool Edit include their own dynamics processor
(which = compressor, among other things); this should give a very good idea of
the sort of parameters needed. The Cool Edit Pro dynamics processor offers the
choice of RMS or Peak envelope detection.

Also, there is a code example (in C) for a hard-knee compressor in the musicdsp
archive:

http://www.musicdsp.org/archive.php


Note that in compressors for music, there is a standard technique to delay the
direct signal by at least half the size of the window used to extract the
envelope so that an incoming peak (e.g. a drum hit or plosive speech consonant)
is wholly reduced, rather than  having the compressor react late, so to speak. A
typical envelope window may only be 10msecs or so, which is not particularly
disturbing in real-time streaming.



Richard Dobson

robert bristow-johnson wrote:

> in article 2f592dd3.0503300144.7544fb03@posting.google.com, EC-AKD at
> akd_ecc@yahoo.com wrote on 03/30/2005 04:44:
> 
> 
>>As I am doing it all alone, I feel lost when I dont see it going
>>anywhere. Suggestions and help from usenet would be really
>>appreciated.
> 
> 
> can't help much at a distance.  i can only suggest, if you have the sound
> recorded on some kind of wave editor (like Cool Edit), that you select the
> quiet pieces and manually boost them to what sounds right, then look at your
> measured loudness and the dB boost for that section and see if you can
> construct a heuristic curve that maps dB loudness (relative to fullscale) to
> dB of boost.
> 
> rots o' ruk.
>