[e2e] Re: crippled Internet

[Christian Huitema]

> From: David P. Reed [mailto:dpreed at reed.com]
> Sent: Thursday, April 26, 2001 8:13 PM
> To: cannara at attglobal.net
> Cc: end2end-interest at postel.org
> Subject: Re: [e2e] Re: crippled Internet
> 
> At 08:36 PM 4/25/01 -0700, Cannara wrote:
> >There are ITU specs for jitter and delay in voice that have been used
for
> >years in standard telco system design.  I should think these would be
> easily
> >accessible, but I only have a couple of graphs on paper.  The basic
idea
> is
> >that 50mS or so of frequent variation in sample arrival time is hard
for
> >listeners, and 100mS roundtrip delay becomes annoying in
conversations.
> 
> 
> This is true, but small numbers of delayed voice frames can be treated
as
> dropped and turned into "interpolated noise bursts" or error corrected
by
> the time they need to be played, as long as the frames around them are
> available.  Thus 50 ms of variation can be achieved by treating frames
> that
> arrive late as dropped.

Well, we may have to think about correlation. Suppose that frames are
sent every 20 or 30 ms. Then suppose that for some reason frame number X
is unacceptably delayed, say by more than 150 ms. There are pretty good
chances that the next frame will be delivered in order, i.e. with a
delay of at least 120-130 ms. In fact, this is the kind of data one
would expect from measurements: what is the correlation between the
delay of successive frames, do we converge rapidly to a better situation
or don't we?

Also, we must realize that dropping frames has a nasty effect on
quality. Many codec rely on long term estimates to achieve better
compression; loosing frames has the effect of "polluting" this long term
estimates, creating noise on following frames; the long term estimates
will only re-converge after a long time. Indeed, the effect can be cured
if the frame is merely delayed, but that is more complex than "treat
them as dropped." In any case, repeated losses also have an effect on
the perceived sound quality. Sure, loss concealment techniques help.
However, experience show that the "user satisfaction" is affected by
losses, and that the effect is a function of both the loss rate and the
loss distribution. Even a low rate error, say 1 or 2%, can be perceived
as nasty if the errors arrive in bursts.

-- Christian Huitema