Subj : Re: Real Modem on VOIP?
To   : Alastair Frankland
From : Grzegorz Worona
Date : Wed Sep 16 2026 09:12 pm

Hi Alastair,

 AF> I'm looking forward to our switch over next year, although it's
 AF> going to put my modems out on a limb and looking at it as a steep
 AF> learning curve.

I cannot offer you a year of hard-won wisdom - my board is a month old
and the modem side of it is younger still. But I am already on the far
side of the switch you are facing: a US Robotics Courier V.Everything
answering behind a Cisco SPA122 ATA, on a Polish VoIP provider, with no
copper anywhere in the path. So let me hand over what those few weeks
cost me to learn, in case it saves you an evening or two.

The ATA settings that actually matter, in rough order of how much pain
they save:

  Modem Line              yes     one switch, turns off echo canceller,
                                  call waiting and silence suppression
  Preferred Codec         G711a   (G711u where you are)
  Use Pref Codec Only     yes     stops mid-call renegotiation
  Silence Supp Enable     no      VAD hears a carrier as "silence"
  Echo Canc / Echo Supp   no      both destroy a data carrier
  Network Jitter Level    extremely high
  Jitter Buffer Adjust    disable every re-adjust is a clock jump,
                                  and a clock jump is a retrain
  RTP Packet Size         0.010   shorter packets, smaller damage
                                  from a single lost one
  FAX / T.38              no      T.38 is for fax, it will not carry
                                  a data modem

A fixed jitter buffer sounds wrong until you think about it: latency
does not bother a modem at all, but *changing* latency does.

Two hardware traps I experienced that had nothing to do with VoIP and
cost me the most time:

  - the ATA needs a 5V/2A supply. An identical-looking 1A brick makes
    the box reset exactly when a call arrives, i.e. when the FXS port
    draws ring current. Looks like a VoIP problem. Is not.

  - on the Courier, DIP switch 10 must be OFF. With it ON the modem
    loads the ROM template at every power-up and silently ignores
    NVRAM, so every AT&W you carefully wrote is gone on the next
    reset. I lost an evening to that one before I compared ATI4
    against ATI5 and saw they disagreed.

On the modem side I also set S10=20 (tolerate 2s of carrier loss
instead of 0.7s) and S7=60, because SIP takes its time to set a call
up. The speed ceiling I left wide open - V.34 fallback works downwards
on its own, and closing it in advance only removes options.

One more thing I am in the middle of, in case anyone here has an
opinion on it. I am swapping the Courier's firmware. The stock USR
images were tuned for the kind of copper you get in North America, and
there is a well-known unofficial line of firmware - ID_SDL - that came
out of Eastern Europe and was calibrated against the local phone plant,
which was a great deal less forgiving. My working theory is that a
modem taught to be stubborn on a bad analogue line may also be more
stubborn on a VoIP path, because the two punish it in similar ways:
timing that wanders, and gaps that should not be there at all.

I have no results yet - that is precisely the experiment. Same line,
same ATA settings, only the firmware changes, and I will report what
both do. If anyone here has already run ID_SDL over VoIP I would very
much like to hear it before I draw conclusions from one modem.

Now the honest part, because I would rather be useful than impressive:
I can prove the *path* is clean, but I cannot yet prove the *result*.
I capture every call on the router and analyse it, and a typical one
looks like this - a 107 second call, both directions:

  packets lost          0 of 10691 out, 0 of 5345 in
  jitter (95th pct)     1.19 ms out, 0.35 ms in
  clock drift           about 14 ppm
  silence before first
  RTP packet            88 ms

So the network is not the enemy here. But every call so far has been me
dialling in from a mobile phone to listen to the answer tone - I have
not yet had a real modem call the line, which means I have a measured
path and no confirmed CONNECT to brag about. The spectrum of those
calls shows exactly that: my Courier sends its answer tone, waits
about 25 seconds for a handshake that is never coming, and falls back.

Greetings from Warsaw,
Grzegorz Worona (kombos)

--- ENiGMA 1/2 v0.5.1-beta (linux; x64)
 * Origin: DEAD SOCKET CHURCH (2:484/101)

.