Digital Mixer.

Neil-

> Last year I went to SPIE's AeroSense
conference and learned about
> Independent Component Analysis and an application called the 'Cocktail
> Party.'  Basically, it's a method to assume that n multiple
signals are
> recorded by n microphones.  ICA is the method that allows the signals to be
> separated into n discrete signals.  Perhaps a look down this path may be
> what you need...

ICA and other techniques aimed at separating source signals is intended to
locate
direction of the signals and also to deliberately focus on signal vs. another. 
At a
cocktail party, you want to listen to the person talking to you and tune out
everyone
else talking around you.  In a conferencing situation, you would not normally
make an
assumption about one listener's preference vs. another -- several people
are in the
room listening to the same speaker.  Whoever is speaking has to be heard by all.
 If
two people are speaking at the same time, then the system would not
automatically
make a choice as to who to emphasize.  Or if you gave users that option, then it
would be a very sophisticated conference system indeed.

Jeff Brower
DSP sw/hw engineer
Signalogic

> -----Original Message-----
> From: Jeff Brower [mailto:jbrower@jbro...]
> Sent: Tuesday, August 20, 2002 7:21 AM
> To: Ulrich Prakash
> Cc: audiodsp@audi...
> Subject: Re: [audiodsp] Digital Mixer.
> 
> Ulrich-
> 
> > In that case,if I need to mix 'm' voice channels(in the form
of RTP
> > packets),then
> > y(n) = a1*x1 + a2*x2 + ... am*xm
> >
> >         = SIGMA ai*xi,where i = 1 to m   (SIGMA - i don't have
SIGMA
> symbol
> > in my mail editor)
> >
> > And SIGMA ai = 1,where i = 1 to m.
> > In that case,if there are 10 input voice channels,then the amplitude
of
> > original sample x1 in output y will be around x1/10.Will this not
affect
> > the fidelity of x1?
> 
> Ahh, now you not mixing, but conferencing -- you should have said so
> earlier.  You
> need some type of algorithm that dynamically looks for "most energetic
> voice" and
> give that channel more weight.  I've seen some conferencing algorithms
that
> will look
> for most 2 or 3 dominant channels and give them weight.
> 
> Clearly if you just constantly multiply every channel by 0.1 you will not
> hear the
> speaker -- he/she will be averaged with noise/silence.  Not everyone talks
> at once,
> right?  Or at least you hope that your algorithm does not cause them to do
> that :-)
> 
> -Jeff
> 
> 
> > At 08:20 AM 8/19/02 -0500, Jeff Brower wrote:
> > >Ulrich Prakash-
> > >
> > > > >Mix == add.  Add your signals together, like this:
> > > > >
> > > > >   y[n] = a*x1[n] + b*x2[n]
> > > > >
> > > > >Suggest that you maintain a + b = 1.
> > > >
> > > > [Prakash]
> > > > Can I take a = 0.5 and b = 0.5,or is there any particular
factor to
> > > > evaluate the value of a and b?
> > > >
> > > > >   Then put a dial on your MATLAB GUI that allows
> > > > >the user to adjust between a and b.  All the way to
left, a = 1.  All
> the
> > > > >way to
> > > > >right, b = 1.
> > > >
> > > > [Prakash]
> > > >  From what you have said,I assume,if
> > > > x1 = {4,2,6} and x2 = {8,10,12},and a=0.5,b=0.5then
> > > > y = {2+4, 1+5, 3+6} = {6, 6, 9}
> > > > Is that so straight-forward?
> > >
> > >Yep.
> > >
> > > >Will I hear both samples x1 and x2 from y output?
> > >
> > >Yep.
> > >
> > > > But I didn't understand what you mean to the LEFT and
RIGHT here.
> > >
> > >When the user turns the dial.  You can put a dial (knob) on your
GUI to
> easily
> > >demonstrate the effect; better than asking your users to enter
values of
> a
> > >and b.
> > >
> > >Jeff Brower
> > >DSP sw/hw engineer
> > >Signalogic

Hello, I just subsribe this list:)
I have a related problem: suppose I have 2 ulaw packet, what is the 
algorithm to mix them and create a new ulaw packet. When remote 
client receive and decode it, he can hear both audio. 

Is it possible?

Kun
	--- In audiodsp@y..., "VandeVoorde, Neil" <nvandevoorde@p...>
wrote:
> Last year I went to SPIE's AeroSense
conference and learned about
> Independent Component Analysis and an application called the 
'Cocktail
> Party.'  Basically, it's a method to
assume that n multiple signals 
are
> recorded by n microphones.  ICA is the method that
allows the 
signals to be
> separated into n discrete signals.  Perhaps a look
down this path 
may be
> what you need...
> 
> Regards,/s/neil 
> 
> Neil E. Van de Voorde, Ph.D.
> Senior Scientist
> Planning Systems Inc.
> 228.689.8775
>  
> 
> -----Original Message-----
> From: Jeff Brower [mailto:jbrower@s...] 
> Sent: Tuesday, August 20, 2002 7:21 AM
> To: Ulrich Prakash
> Cc: audiodsp@y...
> Subject: Re: [audiodsp] Digital Mixer.
> 
> Ulrich-
> 
> > In that case,if I need to mix 'm' voice channels(in the form
of 
RTP
> > packets),then
> > y(n) = a1*x1 + a2*x2 + ... am*xm
> > 
> >         = SIGMA ai*xi,where i = 1 to m   (SIGMA - i don't have 
SIGMA
> symbol
> > in my mail editor)
> > 
> > And SIGMA ai = 1,where i = 1 to m.
> > In that case,if there are 10 input voice channels,then the 
amplitude of
> > original sample x1 in output y will be around
x1/10.Will this not 
affect
> > the fidelity of x1?
> 
> Ahh, now you not mixing, but conferencing -- you should have said so
> earlier.  You
> need some type of algorithm that dynamically looks for "most 
energetic
> voice" and
> give that channel more weight.  I've seen some conferencing 
algorithms that
> will look
> for most 2 or 3 dominant channels and give them weight.
> 
> Clearly if you just constantly multiply every channel by 0.1 you 
will not
> hear the
> speaker -- he/she will be averaged with noise/silence.  Not everyone 
talks
> at once,
> right?  Or at least you hope that your algorithm does not cause them 
to do
> that :-)
> 
> -Jeff
>  
> 
> > At 08:20 AM 8/19/02 -0500, Jeff Brower wrote:
> > >Ulrich Prakash-
> > >
> > > > >Mix == add.  Add your signals together, like this:
> > > > >
> > > > >   y[n] = a*x1[n] + b*x2[n]
> > > > >
> > > > >Suggest that you maintain a + b = 1.
> > > >
> > > > [Prakash]
> > > > Can I take a = 0.5 and b = 0.5,or is there any particular 
factor to
> > > > evaluate the value of a and b?
> > > >
> > > > >   Then put a dial on your MATLAB GUI that allows
> > > > >the user to adjust between a and b.  All the way to
left, a 1.  All
> the
> > > > >way to
> > > > >right, b = 1.
> > > >
> > > > [Prakash]
> > > >  From what you have said,I assume,if
> > > > x1 = {4,2,6} and x2 = {8,10,12},and a=0.5,b=0.5then
> > > > y = {2+4, 1+5, 3+6} = {6, 6, 9}
> > > > Is that so straight-forward?
> > >
> > >Yep.
> > >
> > > >Will I hear both samples x1 and x2 from y output?
> > >
> > >Yep.
> > >
> > > > But I didn't understand what you mean to the LEFT and
RIGHT 
here.
> > >
> > >When the user turns the dial.  You can put a dial (knob) on your 
GUI to
> easily
> > >demonstrate the effect; better than asking your users to enter 
values of
> a
> > >and b.
> > >
> > >Jeff Brower
> > >DSP sw/hw engineer
> > >Signalogic
> 
> 
> _____________________________________
> Note: If you do a simple "reply" with your email client, only the

author of
> this message will receive your answer.  You need
to do a "reply all" 
if you
> want your answer to be distributed to the entire
group.
> 
> _____________________________________
> About this discussion group:
> 
> To Join:  audiodsp-subscribe@y...
> 
> To Post:  audiodsp@y...
> 
> To Leave: audiodsp-unsubscribe@y...
> 
> Archives: http://groups.yahoo.com/group/audiodsp
> 
> Other DSP-Related Groups: http://www.dsprelated.com
>  
> 
> ">http://docs.yahoo.com/info/terms/
	

Kun-

> I have a related problem: suppose I have 2 ulaw
packet, what is the
> algorithm to mix them and create a new ulaw packet. When remote
> client receive and decode it, he can hear both audio.

   y[n] = U(a*u(x1[n]) + b*u(x2[n]))

where

  U(x) = uLaw(x)          (compress)
  u(x) = inverse uLaw(x)  (expand)
  a + b = 1
  y[n] is packet data

Jeff Brower
DSP sw/hw engineer
Signalogic
	> --- In audiodsp@y..., "VandeVoorde, Neil"
<nvandevoorde@p...> wrote:
> > Last year I went to SPIE's AeroSense
conference and learned about
> > Independent Component Analysis and an application called the
> 'Cocktail
> > Party.'  Basically, it's a method to assume that n multiple
signals
> are
> > recorded by n microphones.  ICA is the method that allows the
> signals to be
> > separated into n discrete signals.  Perhaps a look down this path
> may be
> > what you need...
> >
> > Regards,/s/neil
> >
> > Neil E. Van de Voorde, Ph.D.
> > Senior Scientist
> > Planning Systems Inc.
> > 228.689.8775
> >
> >
> > -----Original Message-----
> > From: Jeff Brower [mailto:jbrower@s...]
> > Sent: Tuesday, August 20, 2002 7:21 AM
> > To: Ulrich Prakash
> > Cc: audiodsp@y...
> > Subject: Re: [audiodsp] Digital Mixer.
> >
> > Ulrich-
> >
> > > In that case,if I need to mix 'm' voice channels(in the
form of
> RTP
> > > packets),then
> > > y(n) = a1*x1 + a2*x2 + ... am*xm
> > >
> > >         = SIGMA ai*xi,where i = 1 to m   (SIGMA - i don't
have
> SIGMA
> > symbol
> > > in my mail editor)
> > >
> > > And SIGMA ai = 1,where i = 1 to m.
> > > In that case,if there are 10 input voice channels,then the
> amplitude of
> > > original sample x1 in output y will be around x1/10.Will this not
> affect
> > > the fidelity of x1?
> >
> > Ahh, now you not mixing, but conferencing -- you should have said so
> > earlier.  You
> > need some type of algorithm that dynamically looks for "most
> energetic
> > voice" and
> > give that channel more weight.  I've seen some conferencing
> algorithms that
> > will look
> > for most 2 or 3 dominant channels and give them weight.
> >
> > Clearly if you just constantly multiply every channel by 0.1 you
> will not
> > hear the
> > speaker -- he/she will be averaged with noise/silence.  Not everyone
> talks
> > at once,
> > right?  Or at least you hope that your algorithm does not cause them
> to do
> > that :-)
> >
> > -Jeff
> >
> >
> > > At 08:20 AM 8/19/02 -0500, Jeff Brower wrote:
> > > >Ulrich Prakash-
> > > >
> > > > > >Mix == add.  Add your signals together, like this:
> > > > > >
> > > > > >   y[n] = a*x1[n] + b*x2[n]
> > > > > >
> > > > > >Suggest that you maintain a + b = 1.
> > > > >
> > > > > [Prakash]
> > > > > Can I take a = 0.5 and b = 0.5,or is there any
particular
> factor to
> > > > > evaluate the value of a and b?
> > > > >
> > > > > >   Then put a dial on your MATLAB GUI that allows
> > > > > >the user to adjust between a and b.  All the way to
left, a > 1.  All
> > the
> > > > > >way to
> > > > > >right, b = 1.
> > > > >
> > > > > [Prakash]
> > > > >  From what you have said,I assume,if
> > > > > x1 = {4,2,6} and x2 = {8,10,12},and a=0.5,b=0.5then
> > > > > y = {2+4, 1+5, 3+6} = {6, 6, 9}
> > > > > Is that so straight-forward?
> > > >
> > > >Yep.
> > > >
> > > > >Will I hear both samples x1 and x2 from y output?
> > > >
> > > >Yep.
> > > >
> > > > > But I didn't understand what you mean to the LEFT
and RIGHT
> here.
> > > >
> > > >When the user turns the dial.  You can put a dial (knob) on
your
> GUI to
> > easily
> > > >demonstrate the effect; better than asking your users to
enter
> values of
> > a
> > > >and b.
> > > >
> > > >Jeff Brower
> > > >DSP sw/hw engineer
> > > >Signalogic

Kun-

> Thanks. Where can I find this kind of algorithms
as mentioned below to
> dynamically conferencing?

Not sure really.  We've done some conferencing work before, we just worried
about the
2 channels with most energy at any one time.  3 channels worked Ok, too, but
beyond
that, I seem to recall there was not much to be gained -- if 4 people are
speaking at
once, well, what can you do...it's a mess "in real life", too. 
Also you have to be
careful how you transition the weights -- you can't have a,b,c
"jerk" to different
values; use some type of lowpass/averaging filter on the weights, something that
transitions smoothly, over about 50 to 100 msec.

Jeff Brower
DSP sw/hw engineer
Signalogic
	> > Ahh, now you not mixing, but conferencing -- you should have said so
> > earlier. You
> > need some type of algorithm that dynamically looks for "most
energetic
> > voice" and
> > give that channel more weight. I've seen some conferencing
algorithms that
> > will look
> > for most 2 or 3 dominant channels and give them weight.
> 
> Jeff Brower wrote:
> >
> > Kun-
> >
> > > I have a related problem: suppose I have 2 ulaw packet, what is
the
> > > algorithm to mix them and create a new ulaw packet. When remote
> > > client receive and decode it, he can hear both audio.
> >
> >    y[n] = U(a*u(x1[n]) + b*u(x2[n]))
> >
> > where
> >
> >   U(x) = uLaw(x)          (compress)
> >   u(x) = inverse uLaw(x)  (expand)
> >   a + b = 1
> >   y[n] is packet data
> >
> > Jeff Brower
> > DSP sw/hw engineer
> > Signalogic
> >
> > > --- In audiodsp@y..., "VandeVoorde, Neil"
<nvandevoorde@p...> wrote:
> > > > Last year I went to SPIE's AeroSense conference and
learned about
> > > > Independent Component Analysis and an application called the
> > > 'Cocktail
> > > > Party.'  Basically, it's a method to assume that n
multiple signals
> > > are
> > > > recorded by n microphones.  ICA is the method that allows
the
> > > signals to be
> > > > separated into n discrete signals.  Perhaps a look down this
path
> > > may be
> > > > what you need...
> > > >
> > > > Regards,/s/neil
> > > >
> > > > Neil E. Van de Voorde, Ph.D.
> > > > Senior Scientist
> > > > Planning Systems Inc.
> > > > 228.689.8775
> > > >
> > > >
> > > > -----Original Message-----
> > > > From: Jeff Brower [mailto:jbrower@s...]
> > > > Sent: Tuesday, August 20, 2002 7:21 AM
> > > > To: Ulrich Prakash
> > > > Cc: audiodsp@y...
> > > > Subject: Re: [audiodsp] Digital Mixer.
> > > >
> > > > Ulrich-
> > > >
> > > > > In that case,if I need to mix 'm' voice
channels(in the form of
> > > RTP
> > > > > packets),then
> > > > > y(n) = a1*x1 + a2*x2 + ... am*xm
> > > > >
> > > > >         = SIGMA ai*xi,where i = 1 to m   (SIGMA - i
don't have
> > > SIGMA
> > > > symbol
> > > > > in my mail editor)
> > > > >
> > > > > And SIGMA ai = 1,where i = 1 to m.
> > > > > In that case,if there are 10 input voice channels,then
the
> > > amplitude of
> > > > > original sample x1 in output y will be around
x1/10.Will this not
> > > affect
> > > > > the fidelity of x1?
> > > >
> > > > Ahh, now you not mixing, but conferencing -- you should have
said so
> > > > earlier.  You
> > > > need some type of algorithm that dynamically looks for
"most
> > > energetic
> > > > voice" and
> > > > give that channel more weight.  I've seen some
conferencing
> > > algorithms that
> > > > will look
> > > > for most 2 or 3 dominant channels and give them weight.
> > > >
> > > > Clearly if you just constantly multiply every channel by 0.1
you
> > > will not
> > > > hear the
> > > > speaker -- he/she will be averaged with noise/silence.  Not
everyone
> > > talks
> > > > at once,
> > > > right?  Or at least you hope that your algorithm does not
cause them
> > > to do
> > > > that :-)
> > > >
> > > > -Jeff
> > > >
> > > >
> > > > > At 08:20 AM 8/19/02 -0500, Jeff Brower wrote:
> > > > > >Ulrich Prakash-
> > > > > >
> > > > > > > >Mix == add.  Add your signals together,
like this:
> > > > > > > >
> > > > > > > >   y[n] = a*x1[n] + b*x2[n]
> > > > > > > >
> > > > > > > >Suggest that you maintain a + b = 1.
> > > > > > >
> > > > > > > [Prakash]
> > > > > > > Can I take a = 0.5 and b = 0.5,or is there
any particular
> > > factor to
> > > > > > > evaluate the value of a and b?
> > > > > > >
> > > > > > > >   Then put a dial on your MATLAB GUI
that allows
> > > > > > > >the user to adjust between a and b.  All
the way to left, a > > > 1.  All
> > > > the
> > > > > > > >way to
> > > > > > > >right, b = 1.
> > > > > > >
> > > > > > > [Prakash]
> > > > > > >  From what you have said,I assume,if
> > > > > > > x1 = {4,2,6} and x2 = {8,10,12},and
a=0.5,b=0.5then
> > > > > > > y = {2+4, 1+5, 3+6} = {6, 6, 9}
> > > > > > > Is that so straight-forward?
> > > > > >
> > > > > >Yep.
> > > > > >
> > > > > > >Will I hear both samples x1 and x2 from y
output?
> > > > > >
> > > > > >Yep.
> > > > > >
> > > > > > > But I didn't understand what you mean to
the LEFT and RIGHT
> > > here.
> > > > > >
> > > > > >When the user turns the dial.  You can put a dial
(knob) on your
> > > GUI to
> > > > easily
> > > > > >demonstrate the effect; better than asking your
users to enter
> > > values of
> > > > a
> > > > > >and b.
> > > > > >
> > > > > >Jeff Brower
> > > > > >DSP sw/hw engineer
> > > > > >Signalogic

Hi,

Using the summing algorithm,I feel that we are not making MAXIMUM 
utilization of the mixed Output Channel,if I got it right.

That is,though we take a + b + c = 1(for 3 inputs),we have to subtract one 
input with coefficient(i.e. same sample)before sending.Hence,b+c<1 and
NOT 
equal to 1,to have maximum utilization.
	Let me explain the scenario more in detail:

Consider there are 3 users(A,B,C) in a conference and there is ONE 
conference server which mixes the RTP packets.
For 3 inputs,the mixed output,
  y = ax1+bx2+cx3,   ------(1)
and a+b+c=1.            ---(2)

However,we CANNOT send output 'y' as it is,to A,B and C.
This is because,A should not hear his own voice,....likewise for B,C also.

Hence we have to recalculate 3 outputs,like
y1 = y - ax1
y2 = y - bx2  --------------(3)
y3 = y - cx3
and then send y1(put into RTP packet) to A,y2 to B and y3 to C.

If this is the case,
y1 = y- ax1 = bx2+cx3 and here b+c <1.
Hence here actually the outputs coefficients b+c<1,and NOT EQUAL to 1 and

hence we are not making maximum utilization of the output channel.

ALTERNATIVE solution:

The alternative is doing a calculation like this:
y1 = b1*x2 + c1*x3 , where b1 + c1 = 1,
y2 = a2*x1 + c2*x3 , where a2 + c2 = 1,
y3 = a3*x1 + b3*x2 , where a3 + b3 = 1,

But this method will involve MORE COMPUTATION TIME due to more no. of 
multiplications and additions and finding co-efficients,if we mix more than 
3 channels.
	So must one follow this ALTERNATIVE solution with more computations or the 
PREVIOUS one with lesser utilization of the channel?

Also please let me know if there is some other way of dealing this situation.

Thanks in advance,
Prakash.

At 11:44 PM 8/21/02 -0500, Jeff Brower wrote:
>Kun-
>
> > Thanks. Where can I find this kind of algorithms as mentioned below to
> > dynamically conferencing?
>
>Not sure really.  We've done some conferencing work before, we just 
>worried about the
>2 channels with most energy at any one time.  3 channels worked Ok, too, 
>but beyond
>that, I seem to recall there was not much to be gained -- if 4 people are 
>speaking at
>once, well, what can you do...it's a mess "in real life",
too.  Also you 
>have to be
>careful how you transition the weights -- you can't have a,b,c
"jerk" to 
>different
>values; use some type of lowpass/averaging filter on the weights, 
>something that
>transitions smoothly, over about 50 to 100 msec.
>
>Jeff Brower
>DSP sw/hw engineer
>Signalogic
>
>
>
> > > Ahh, now you not mixing, but conferencing -- you should have said
so
> > > earlier. You
> > > need some type of algorithm that dynamically looks for "most
energetic
> > > voice" and
> > > give that channel more weight. I've seen some conferencing
algorithms 
> that
> > > will look
> > > for most 2 or 3 dominant channels and give them weight.
> >
> > Jeff Brower wrote:
> > >
> > > Kun-
> > >
> > > > I have a related problem: suppose I have 2 ulaw packet, what
is the
> > > > algorithm to mix them and create a new ulaw packet. When
remote
> > > > client receive and decode it, he can hear both audio.
> > >
> > >    y[n] = U(a*u(x1[n]) + b*u(x2[n]))
> > >
> > > where
> > >
> > >   U(x) = uLaw(x)          (compress)
> > >   u(x) = inverse uLaw(x)  (expand)
> > >   a + b = 1
> > >   y[n] is packet data
> > >
> > > Jeff Brower
> > > DSP sw/hw engineer
> > > Signalogic
> > >
> > > > --- In audiodsp@y..., "VandeVoorde, Neil"
<nvandevoorde@p...> wrote:
> > > > > Last year I went to SPIE's AeroSense conference
and learned about
> > > > > Independent Component Analysis and an application
called the
> > > > 'Cocktail
> > > > > Party.'  Basically, it's a method to assume
that n multiple signals
> > > > are
> > > > > recorded by n microphones.  ICA is the method that
allows the
> > > > signals to be
> > > > > separated into n discrete signals.  Perhaps a look down
this path
> > > > may be
> > > > > what you need...
> > > > >
> > > > > Regards,/s/neil
> > > > >
> > > > > Neil E. Van de Voorde, Ph.D.
> > > > > Senior Scientist
> > > > > Planning Systems Inc.
> > > > > 228.689.8775
> > > > >
> > > > >
> > > > > -----Original Message-----
> > > > > From: Jeff Brower [mailto:jbrower@s...]
> > > > > Sent: Tuesday, August 20, 2002 7:21 AM
> > > > > To: Ulrich Prakash
> > > > > Cc: audiodsp@y...
> > > > > Subject: Re: [audiodsp] Digital Mixer.
> > > > >
> > > > > Ulrich-
> > > > >
> > > > > > In that case,if I need to mix 'm' voice
channels(in the form of
> > > > RTP
> > > > > > packets),then
> > > > > > y(n) = a1*x1 + a2*x2 + ... am*xm
> > > > > >
> > > > > >         = SIGMA ai*xi,where i = 1 to m   (SIGMA -
i don't have
> > > > SIGMA
> > > > > symbol
> > > > > > in my mail editor)
> > > > > >
> > > > > > And SIGMA ai = 1,where i = 1 to m.
> > > > > > In that case,if there are 10 input voice
channels,then the
> > > > amplitude of
> > > > > > original sample x1 in output y will be around
x1/10.Will this not
> > > > affect
> > > > > > the fidelity of x1?
> > > > >
> > > > > Ahh, now you not mixing, but conferencing -- you should
have said so
> > > > > earlier.  You
> > > > > need some type of algorithm that dynamically looks for
"most
> > > > energetic
> > > > > voice" and
> > > > > give that channel more weight.  I've seen some
conferencing
> > > > algorithms that
> > > > > will look
> > > > > for most 2 or 3 dominant channels and give them weight.
> > > > >
> > > > > Clearly if you just constantly multiply every channel
by 0.1 you
> > > > will not
> > > > > hear the
> > > > > speaker -- he/she will be averaged with noise/silence. 
Not everyone
> > > > talks
> > > > > at once,
> > > > > right?  Or at least you hope that your algorithm does
not cause them
> > > > to do
> > > > > that :-)
> > > > >
> > > > > -Jeff
> > > > >
> > > > >
> > > > > > At 08:20 AM 8/19/02 -0500, Jeff Brower wrote:
> > > > > > >Ulrich Prakash-
> > > > > > >
> > > > > > > > >Mix == add.  Add your signals
together, like this:
> > > > > > > > >
> > > > > > > > >   y[n] = a*x1[n] + b*x2[n]
> > > > > > > > >
> > > > > > > > >Suggest that you maintain a + b = 1.
> > > > > > > >
> > > > > > > > [Prakash]
> > > > > > > > Can I take a = 0.5 and b = 0.5,or is
there any particular
> > > > factor to
> > > > > > > > evaluate the value of a and b?
> > > > > > > >
> > > > > > > > >   Then put a dial on your MATLAB
GUI that allows
> > > > > > > > >the user to adjust between a and b. 
All the way to left, a > > > > 1.  All
> > > > > the
> > > > > > > > >way to
> > > > > > > > >right, b = 1.
> > > > > > > >
> > > > > > > > [Prakash]
> > > > > > > >  From what you have said,I assume,if
> > > > > > > > x1 = {4,2,6} and x2 = {8,10,12},and
a=0.5,b=0.5then
> > > > > > > > y = {2+4, 1+5, 3+6} = {6, 6,
9}
> > > > > > > > Is that so straight-forward?
> > > > > > >
> > > > > > >Yep.
> > > > > > >
> > > > > > > >Will I hear both samples x1 and x2 from y
output?
> > > > > > >
> > > > > > >Yep.
> > > > > > >
> > > > > > > > But I didn't understand what you
mean to the LEFT and RIGHT
> > > > here.
> > > > > > >
> > > > > > >When the user turns the dial.  You can put a
dial (knob) on your
> > > > GUI to
> > > > > easily
> > > > > > >demonstrate the effect; better than asking
your users to enter
> > > > values of
> > > > > a
> > > > > > >and b.
> > > > > > >
> > > > > > >Jeff Brower
> > > > > > >DSP sw/hw engineer
> > > > > > >Signalogic
>
>
>_____________________________________
>Note: If you do a simple "reply" with your email client, only the
author 
>of this message will receive your answer.  You need to do a "reply
all" if 
>you want your answer to be distributed to the entire group.
>
>_____________________________________
>About this discussion group:
>
>To Join:  audiodsp-subscribe@audi...
>
>To Post:  audiodsp@audi...
>
>To Leave: audiodsp-unsubscribe@audi...
>
>Archives: http://groups.yahoo.com/group/audiodsp
>
>Other DSP-Related Groups: http://www.dsprelated.com
>
>
>">http://docs.yahoo.com/info/terms/
	

Ulrich-

As you can see, it's more difficult when the speakers are in different
locations and
the mixed sound comes from a central server.  Do you have control of processing
at
each speaker's location?  If so, you can do something simple (like a
typical Office
Max phone) and mute the sound when a person is speaking -- they can't hear
anything,
so your processing requirements stay the same.  Basically, that becomes a 1-way
sound
situation (also known as half-duplex in communication systems).

If you can't do that, or you have processing ability only at the server,
then you
will need to subtract or selectively mix, as you describe.

You are approaching echo canceller / adaptive signal cancellation techniques,
which
are used to improve conferencing quality and provide a constant 2-way sound
situation
(full-duplex).  But if you do that, you will need a DSP or other fast processor,
and
cannot rely on the PC/sound card to do it.

Jeff Brower
DSP sw/hw engineer
Signalogic
	Ulrich Prakash wrote:
> 
> Hi,
> 
> Using the summing algorithm,I feel that we are not making MAXIMUM
> utilization of the mixed Output Channel,if I got it right.
> 
> That is,though we take a + b + c = 1(for 3 inputs),we have to subtract one
> input with coefficient(i.e. same sample)before sending.Hence,b+c<1
and NOT
> equal to 1,to have maximum utilization.
> 
> Let me explain the scenario more in detail:
> 
> Consider there are 3 users(A,B,C) in a conference and there is ONE
> conference server which mixes the RTP packets.
> For 3 inputs,the mixed output,
>   y = ax1+bx2+cx3,   ------(1)
> and a+b+c=1.            ---(2)
> 
> However,we CANNOT send output 'y' as it is,to A,B and C.
> This is because,A should not hear his own voice,....likewise for B,C also.
> 
> Hence we have to recalculate 3 outputs,like
> y1 = y - ax1
> y2 = y - bx2  --------------(3)
> y3 = y - cx3
> and then send y1(put into RTP packet) to A,y2 to B and y3 to C.
> 
> If this is the case,
> y1 = y- ax1 = bx2+cx3 and here b+c <1.
> Hence here actually the outputs coefficients b+c<1,and NOT EQUAL to
1 and
> hence we are not making maximum utilization of the output channel.
> 
> ALTERNATIVE solution:
> 
> The alternative is doing a calculation like this:
> y1 = b1*x2 + c1*x3 , where b1 + c1 = 1,
> y2 = a2*x1 + c2*x3 , where a2 + c2 = 1,
> y3 = a3*x1 + b3*x2 , where a3 + b3 = 1,
> 
> But this method will involve MORE COMPUTATION TIME due to more no. of
> multiplications and additions and finding co-efficients,if we mix more than
> 3 channels.
> 
> So must one follow this ALTERNATIVE solution with more computations or the
> PREVIOUS one with lesser utilization of the channel?
> 
> Also please let me know if there is some other way of dealing this
situation.
> 
> Thanks in advance,
> Prakash.
> 
> At 11:44 PM 8/21/02 -0500, Jeff Brower wrote:
> >Kun-
> >
> > > Thanks. Where can I find this kind of algorithms as mentioned
below to
> > > dynamically conferencing?
> >
> >Not sure really.  We've done some conferencing work before, we
just
> >worried about the
> >2 channels with most energy at any one time.  3 channels worked Ok,
too,
> >but beyond
> >that, I seem to recall there was not much to be gained -- if 4 people
are
> >speaking at
> >once, well, what can you do...it's a mess "in real
life", too.  Also you
> >have to be
> >careful how you transition the weights -- you can't have a,b,c
"jerk" to
> >different
> >values; use some type of lowpass/averaging filter on the weights,
> >something that
> >transitions smoothly, over about 50 to 100 msec.
> >
> >Jeff Brower
> >DSP sw/hw engineer
> >Signalogic
> >
> >
> >
> > > > Ahh, now you not mixing, but conferencing -- you should have
said so
> > > > earlier. You
> > > > need some type of algorithm that dynamically looks for
"most energetic
> > > > voice" and
> > > > give that channel more weight. I've seen some
conferencing algorithms
> > that
> > > > will look
> > > > for most 2 or 3 dominant channels and give them weight.
> > >
> > > Jeff Brower wrote:
> > > >
> > > > Kun-
> > > >
> > > > > I have a related problem: suppose I have 2 ulaw packet,
what is the
> > > > > algorithm to mix them and create a new ulaw packet.
When remote
> > > > > client receive and decode it, he can hear both audio.
> > > >
> > > >    y[n] = U(a*u(x1[n]) + b*u(x2[n]))
> > > >
> > > > where
> > > >
> > > >   U(x) = uLaw(x)          (compress)
> > > >   u(x) = inverse uLaw(x)  (expand)
> > > >   a + b = 1
> > > >   y[n] is packet data
> > > >
> > > > Jeff Brower
> > > > DSP sw/hw engineer
> > > > Signalogic
> > > >
> > > > > --- In audiodsp@y..., "VandeVoorde, Neil"
<nvandevoorde@p...> wrote:
> > > > > > Last year I went to SPIE's AeroSense
conference and learned about
> > > > > > Independent Component Analysis and an application
called the
> > > > > 'Cocktail
> > > > > > Party.'  Basically, it's a method to
assume that n multiple signals
> > > > > are
> > > > > > recorded by n microphones.  ICA is the method that
allows the
> > > > > signals to be
> > > > > > separated into n discrete signals.  Perhaps a look
down this path
> > > > > may be
> > > > > > what you need...
> > > > > >
> > > > > > Regards,/s/neil
> > > > > >
> > > > > > Neil E. Van de Voorde, Ph.D.
> > > > > > Senior Scientist
> > > > > > Planning Systems Inc.
> > > > > > 228.689.8775
> > > > > >
> > > > > >
> > > > > > -----Original Message-----
> > > > > > From: Jeff Brower [mailto:jbrower@s...]
> > > > > > Sent: Tuesday, August 20, 2002 7:21 AM
> > > > > > To: Ulrich Prakash
> > > > > > Cc: audiodsp@y...
> > > > > > Subject: Re: [audiodsp] Digital Mixer.
> > > > > >
> > > > > > Ulrich-
> > > > > >
> > > > > > > In that case,if I need to mix 'm'
voice channels(in the form of
> > > > > RTP
> > > > > > > packets),then
> > > > > > > y(n) = a1*x1 + a2*x2 + ... am*xm
> > > > > > >
> > > > > > >         = SIGMA ai*xi,where i = 1 to m  
(SIGMA - i don't have
> > > > > SIGMA
> > > > > > symbol
> > > > > > > in my mail editor)
> > > > > > >
> > > > > > > And SIGMA ai = 1,where i = 1 to m.
> > > > > > > In that case,if there are 10 input voice
channels,then the
> > > > > amplitude of
> > > > > > > original sample x1 in output y will be around
x1/10.Will this not
> > > > > affect
> > > > > > > the fidelity of x1?
> > > > > >
> > > > > > Ahh, now you not mixing, but conferencing -- you
should have said so
> > > > > > earlier.  You
> > > > > > need some type of algorithm that dynamically looks
for "most
> > > > > energetic
> > > > > > voice" and
> > > > > > give that channel more weight.  I've seen
some conferencing
> > > > > algorithms that
> > > > > > will look
> > > > > > for most 2 or 3 dominant channels and give them
weight.
> > > > > >
> > > > > > Clearly if you just constantly multiply every
channel by 0.1 you
> > > > > will not
> > > > > > hear the
> > > > > > speaker -- he/she will be averaged with
noise/silence.  Not everyone
> > > > > talks
> > > > > > at once,
> > > > > > right?  Or at least you hope that your algorithm
does not cause them
> > > > > to do
> > > > > > that :-)
> > > > > >
> > > > > > -Jeff
> > > > > >
> > > > > >
> > > > > > > At 08:20 AM 8/19/02 -0500, Jeff Brower wrote:
> > > > > > > >Ulrich Prakash-
> > > > > > > >
> > > > > > > > > >Mix == add.  Add your signals
together, like this:
> > > > > > > > > >
> > > > > > > > > >   y[n] = a*x1[n] + b*x2[n]
> > > > > > > > > >
> > > > > > > > > >Suggest that you maintain a + b
= 1.
> > > > > > > > >
> > > > > > > > > [Prakash]
> > > > > > > > > Can I take a = 0.5 and b = 0.5,or
is there any particular
> > > > > factor to
> > > > > > > > > evaluate the value of a and b?
> > > > > > > > >
> > > > > > > > > >   Then put a dial on your
MATLAB GUI that allows
> > > > > > > > > >the user to adjust between a
and b.  All the way to left, a > > > > > 1.  All
> > > > > > the
> > > > > > > > > >way to
> > > > > > > > > >right, b = 1.
> > > > > > > > >
> > > > > > > > > [Prakash]
> > > > > > > > >  From what you have said,I
assume,if
> > > > > > > > > x1 = {4,2,6} and x2 = {8,10,12},and
a=0.5,b=0.5then
> > > > > > > > > y = {2+4, 1+5, 3+6} =
{6, 6, 9}
> > > > > > > > > Is that so straight-forward?
> > > > > > > >
> > > > > > > >Yep.
> > > > > > > >
> > > > > > > > >Will I hear both samples x1 and x2
from y output?
> > > > > > > >
> > > > > > > >Yep.
> > > > > > > >
> > > > > > > > > But I didn't understand what
you mean to the LEFT and RIGHT
> > > > > here.
> > > > > > > >
> > > > > > > >When the user turns the dial.  You can
put a dial (knob) on your
> > > > > GUI to
> > > > > > easily
> > > > > > > >demonstrate the effect; better than
asking your users to enter
> > > > > values of
> > > > > > a
> > > > > > > >and b.
> > > > > > > >
> > > > > > > >Jeff Brower
> > > > > > > >DSP sw/hw engineer
> > > > > > > >Signalogic

Jeff Brower wrote:
> 
> You are approaching echo canceller / adaptive signal cancellation
techniques, which
> are used to improve conferencing quality and provide a constant 2-way sound
situation
> (full-duplex).  But if you do that, you will need a DSP or other fast
processor, and
> cannot rely on the PC/sound card to do it.

Well, high quality echo cancellation on PC platforms _is_
possible.

Regards,
Alexander

-- 
dipl. ing.
alexander lerch

zplane.development
http://www.zplane.de
holsteinische str. 39-42
D-12161 berlin
 fon: +49.30.854 09 15.0
 fax: +49.30.854 09 15.5
	

I have tried A+B, A+B/2, it seems worse. I end up using algorithm 
A+B-A*B/max (A and B are 2 different audio sample from 0 to max)

Kun

Jeff Brower wrote:
> 
> Ulrich-
> 
> As you can see, it's more difficult when the speakers are in different
locations and
> the mixed sound comes from a central server.  Do you have control of
processing at
> each speaker's location?  If so, you can do something simple (like a
typical Office
> Max phone) and mute the sound when a person is speaking -- they can't
hear anything,
> so your processing requirements stay the same.  Basically, that becomes a
1-way sound
> situation (also known as half-duplex in communication systems).
> 
> If you can't do that, or you have processing ability only at the
server, then you
> will need to subtract or selectively mix, as you describe.
> 
> You are approaching echo canceller / adaptive signal cancellation
techniques, which
> are used to improve conferencing quality and provide a constant 2-way sound
situation
> (full-duplex).  But if you do that, you will need a DSP or other fast
processor, and
> cannot rely on the PC/sound card to do it.
> 
> Jeff Brower
> DSP sw/hw engineer
> Signalogic
> 
> Ulrich Prakash wrote:
> >
> > Hi,
> >
> > Using the summing algorithm,I feel that we are not making MAXIMUM
> > utilization of the mixed Output Channel,if I got it right.
> >
> > That is,though we take a + b + c = 1(for 3 inputs),we have to subtract
one
> > input with coefficient(i.e. same sample)before
sending.Hence,b+c<1 and NOT
> > equal to 1,to have maximum utilization.
> >
> > Let me explain the scenario more in detail:
> >
> > Consider there are 3 users(A,B,C) in a conference and there is ONE
> > conference server which mixes the RTP packets.
> > For 3 inputs,the mixed output,
> >   y = ax1+bx2+cx3,   ------(1)
> > and a+b+c=1.            ---(2)
> >
> > However,we CANNOT send output 'y' as it is,to A,B and C.
> > This is because,A should not hear his own voice,....likewise for B,C
also.
> >
> > Hence we have to recalculate 3 outputs,like
> > y1 = y - ax1
> > y2 = y - bx2  --------------(3)
> > y3 = y - cx3
> > and then send y1(put into RTP packet) to A,y2 to B and y3 to C.
> >
> > If this is the case,
> > y1 = y- ax1 = bx2+cx3 and here b+c <1.
> > Hence here actually the outputs coefficients b+c<1,and NOT
EQUAL to 1 and
> > hence we are not making maximum utilization of the output channel.
> >
> > ALTERNATIVE solution:
> > 
> > The alternative is doing a calculation like this:
> > y1 = b1*x2 + c1*x3 , where b1 + c1 = 1,
> > y2 = a2*x1 + c2*x3 , where a2 + c2 = 1,
> > y3 = a3*x1 + b3*x2 , where a3 + b3 = 1,
> >
> > But this method will involve MORE COMPUTATION TIME due to more no. of
> > multiplications and additions and finding co-efficients,if we mix more
than
> > 3 channels.
> >
> > So must one follow this ALTERNATIVE solution with more computations or
the
> > PREVIOUS one with lesser utilization of the channel?
> >
> > Also please let me know if there is some other way of dealing this
situation.
> >
> > Thanks in advance,
> > Prakash.
> >
> > At 11:44 PM 8/21/02 -0500, Jeff Brower wrote:
> > >Kun-
> > >
> > > > Thanks. Where can I find this kind of algorithms as
mentioned below to
> > > > dynamically conferencing?
> > >
> > >Not sure really.  We've done some conferencing work before,
we just
> > >worried about the
> > >2 channels with most energy at any one time.  3 channels worked
Ok, too,
> > >but beyond
> > >that, I seem to recall there was not much to be gained -- if 4
people are
> > >speaking at
> > >once, well, what can you do...it's a mess "in real
life", too.  Also you
> > >have to be
> > >careful how you transition the weights -- you can't have
a,b,c "jerk" to
> > >different
> > >values; use some type of lowpass/averaging filter on the weights,
> > >something that
> > >transitions smoothly, over about 50 to 100 msec.
> > >
> > >Jeff Brower
> > >DSP sw/hw engineer
> > >Signalogic
> > >
> > >
> > >
> > > > > Ahh, now you not mixing, but conferencing -- you should
have said so
> > > > > earlier. You
> > > > > need some type of algorithm that dynamically looks for
"most energetic
> > > > > voice" and
> > > > > give that channel more weight. I've seen some
conferencing algorithms
> > > that
> > > > > will look
> > > > > for most 2 or 3 dominant channels and give them weight.
> > > >
> > > > Jeff Brower wrote:
> > > > >
> > > > > Kun-
> > > > >
> > > > > > I have a related problem: suppose I have 2 ulaw
packet, what is the
> > > > > > algorithm to mix them and create a new ulaw
packet. When remote
> > > > > > client receive and decode it, he can hear both
audio.
> > > > >
> > > > >    y[n] = U(a*u(x1[n]) + b*u(x2[n]))
> > > > >
> > > > > where
> > > > >
> > > > >   U(x) = uLaw(x)          (compress)
> > > > >   u(x) = inverse uLaw(x)  (expand)
> > > > >   a + b = 1
> > > > >   y[n] is packet data
> > > > >
> > > > > Jeff Brower
> > > > > DSP sw/hw engineer
> > > > > Signalogic
> > > > >
> > > > > > --- In audiodsp@y..., "VandeVoorde,
Neil" <nvandevoorde@p...> wrote:
> > > > > > > Last year I went to SPIE's AeroSense
conference and learned about
> > > > > > > Independent Component Analysis and an
application called the
> > > > > > 'Cocktail
> > > > > > > Party.'  Basically, it's a method
to assume that n multiple signals
> > > > > > are
> > > > > > > recorded by n microphones.  ICA is the method
that allows the
> > > > > > signals to be
> > > > > > > separated into n discrete signals.  Perhaps a
look down this path
> > > > > > may be
> > > > > > > what you need...
> > > > > > >
> > > > > > > Regards,/s/neil
> > > > > > >
> > > > > > > Neil E. Van de Voorde, Ph.D.
> > > > > > > Senior Scientist
> > > > > > > Planning Systems Inc.
> > > > > > > 228.689.8775
> > > > > > >
> > > > > > >
> > > > > > > -----Original Message-----
> > > > > > > From: Jeff Brower [mailto:jbrower@s...]
> > > > > > > Sent: Tuesday, August 20, 2002 7:21 AM
> > > > > > > To: Ulrich Prakash
> > > > > > > Cc: audiodsp@y...
> > > > > > > Subject: Re: [audiodsp] Digital Mixer.
> > > > > > >
> > > > > > > Ulrich-
> > > > > > >
> > > > > > > > In that case,if I need to mix
'm' voice channels(in the form of
> > > > > > RTP
> > > > > > > > packets),then
> > > > > > > > y(n) = a1*x1 + a2*x2 + ... am*xm
> > > > > > > >
> > > > > > > >         = SIGMA ai*xi,where i = 1 to m  
(SIGMA - i don't have
> > > > > > SIGMA
> > > > > > > symbol
> > > > > > > > in my mail editor)
> > > > > > > >
> > > > > > > > And SIGMA ai = 1,where i = 1 to m.
> > > > > > > > In that case,if there are 10 input voice
channels,then the
> > > > > > amplitude of
> > > > > > > > original sample x1 in output y will be
around x1/10.Will this not
> > > > > > affect
> > > > > > > > the fidelity of x1?
> > > > > > >
> > > > > > > Ahh, now you not mixing, but conferencing --
you should have said so
> > > > > > > earlier.  You
> > > > > > > need some type of algorithm that dynamically
looks for "most
> > > > > > energetic
> > > > > > > voice" and
> > > > > > > give that channel more weight.  I've
seen some conferencing
> > > > > > algorithms that
> > > > > > > will look
> > > > > > > for most 2 or 3 dominant channels and give
them weight.
> > > > > > >
> > > > > > > Clearly if you just constantly multiply every
channel by 0.1 you
> > > > > > will not
> > > > > > > hear the
> > > > > > > speaker -- he/she will be averaged with
noise/silence.  Not everyone
> > > > > > talks
> > > > > > > at once,
> > > > > > > right?  Or at least you hope that your
algorithm does not cause them
> > > > > > to do
> > > > > > > that :-)
> > > > > > >
> > > > > > > -Jeff
> > > > > > >
> > > > > > >
> > > > > > > > At 08:20 AM 8/19/02 -0500, Jeff Brower
wrote:
> > > > > > > > >Ulrich Prakash-
> > > > > > > > >
> > > > > > > > > > >Mix == add.  Add your
signals together, like this:
> > > > > > > > > > >
> > > > > > > > > > >   y[n] = a*x1[n] +
b*x2[n]
> > > > > > > > > > >
> > > > > > > > > > >Suggest that you maintain
a + b = 1.
> > > > > > > > > >
> > > > > > > > > > [Prakash]
> > > > > > > > > > Can I take a = 0.5 and b =
0.5,or is there any particular
> > > > > > factor to
> > > > > > > > > > evaluate the value of a and b?
> > > > > > > > > >
> > > > > > > > > > >   Then put a dial on your
MATLAB GUI that allows
> > > > > > > > > > >the user to adjust between
a and b.  All the way to left, a > > > > > > 1.  All
> > > > > > > the
> > > > > > > > > > >way to
> > > > > > > > > > >right, b = 1.
> > > > > > > > > >
> > > > > > > > > > [Prakash]
> > > > > > > > > >  From what you have said,I
assume,if
> > > > > > > > > > x1 = {4,2,6} and x2 =
{8,10,12},and a=0.5,b=0.5then
> > > > > > > > > > y = {2+4, 1+5,
3+6} = {6, 6, 9}
> > > > > > > > > > Is that so straight-forward?
> > > > > > > > >
> > > > > > > > >Yep.
> > > > > > > > >
> > > > > > > > > >Will I hear both samples x1 and
x2 from y output?
> > > > > > > > >
> > > > > > > > >Yep.
> > > > > > > > >
> > > > > > > > > > But I didn't understand
what you mean to the LEFT and RIGHT
> > > > > > here.
> > > > > > > > >
> > > > > > > > >When the user turns the dial.  You
can put a dial (knob) on your
> > > > > > GUI to
> > > > > > > easily
> > > > > > > > >demonstrate the effect; better than
asking your users to enter
> > > > > > values of
> > > > > > > a
> > > > > > > > >and b.
> > > > > > > > >
> > > > > > > > >Jeff Brower
> > > > > > > > >DSP sw/hw engineer
> > > > > > > > >Signalogic

-- 
-----------------------------
 Kun Wei
 Email:   weikun@weik...
 Address: MC 256-48, Caltech
          Pasadena, CA 91125
 Phone:   1-626-395-8767
-----------------------------
	

Alexander-

> > You are approaching echo canceller / adaptive
signal cancellation techniques, which
> > are used to improve conferencing quality and provide a constant 2-way
sound situation
> > (full-duplex).  But if you do that, you will need a DSP or other fast
processor, and
> > cannot rely on the PC/sound card to do it.
> 
> Well, high quality echo cancellation on PC platforms _is_ possible.

How many channels?  Maybe a few is Ok.  But then you need G.7xx and GSM codecs,
VAD,
jitter buffer, and pretty soon you have a large gateway box that takes forever
to
open a Word document.  I remember well the problems with WinModems, and I
wouldn't go
there again.  Unfortunately it's a Win9x and WinXP world, and doesn't
mix well with
NSP.

Jeff Brower
DSP sw/hw engineer
Signalogic

> --
> dipl. ing.
> alexander lerch
> 
> zplane.development
> http://www.zplane.de
> holsteinische str. 39-42
> D-12161 berlin
>  fon: +49.30.854 09 15.0
>  fax: +49.30.854 09 15.5

Hi Jeff,

Jeff Brower wrote:
> Alexander-
>
>
>>>You are approaching echo canceller / adaptive signal
cancellation techniques, which
>>>are used to improve conferencing quality
and provide a
constant 2-way sound situation
>>>(full-duplex).  But if you do that, you
will need a DSP or
other fast processor, and
>>>cannot rely on the PC/sound card to do it.
>>
>>Well, high quality echo cancellation on PC platforms _is_
possible.
>
>
> How many channels?  Maybe a few is Ok.  But then you need
G.7xx and GSM codecs, VAD,
> jitter buffer, and pretty soon you have a large
gateway box
that takes forever to
> open a Word document.  I remember well the
problems with
WinModems, and I wouldn't go
> there again.  Unfortunately it's a Win9x and
WinXP world,
and doesn't mix well with
> NSP.
>

The number of channels is obviously dependent on the
samplerate and the maximum allowed delay. At 8kHz and 200ms
delay we found that you can run 20-30 (optimized) cancellers
in parallel on a 1GHz-machine.
However, in most cases only one or two cancellers will run
(especially on machines where you can open Word) and you will
have more than enough performance headroom for other
applications. Speech codecs are not as complex as that would
be a problem on a modern PC. But, you are right that a more
complex system should to be designed carefully.

Regards,
Alexander

-- 
dipl. ing.
alexander lerch

zplane.development
http://www.zplane.de
holsteinische str. 39-42
D-12161 berlin
 fon: +49.30.854 09 15.0
 fax: +49.30.854 09 15.5