V.90 digital modem: technical details¶

This is the technical companion to V.90: the digital modem, the plain-English overview. Start there; this page keeps the full detail: the code's names, the Recommendations' clauses, the numbers and the test lists.

Overview¶

dsp/v90.cpp is the digital modem's PCM transmitter. dsp/dp_v34.cpp runs its start-up on top of the V.34 pump.

The modem must run on an 8000 Hz line on a G.711 network: the pump only offers V.90 (and V.92) as the digital modem when the line is at 8000 Hz (v90 in V34Pump::start()). dmodemd runs SIP and PBX lines at that rate (DMODEM_RATE); in the tests the 8000 Hz build of the open modem is half_open8 (compiled with MODEM_RATE=8000).

The pump sends exact codewords, as their linear values, which the network's encoder passes on unchanged. The G.711 law comes from the call's codec; without one, it is µ-law, or A-law with DSP_G711=alaw.

Terms used on this page:

• PCM: here, the G.711 codes sent as exact values, rather than a modulated signal.
• Ucode: a PCM code number from 0 to 127.
• DIL (digital impairment learning): a sequence of known codes that shows the analogue modem what the network does to each code.
• mse: mean square error, a measure of a receiver's noise.

The start-up¶

Phases 1 and 2¶

• In V.8, CM or JM carries PCM availability with the digital modem's bit.
• Phase 2 is the V.34 call modem's part, whichever side called: INFO0d, tone B, probing, INFO1d.
• The analogue modem's INFO1a asks either for V.90 (symbol rate 6, with UINFO, the Ucode of TRN1d's two points) or for V.34.

Phase 3¶

• In: the analogue modem's S, PP, TRN and Ja (Ja carries the DIL descriptor) go into the V.34 receiver.
• Out: Sd, TRN1d, Jd, Jd' and the DIL that Ja asked for go out as PCM. TRN1d lasts 1 s (trn1d_len, see the Agere modem below).
• While it takes in the DIL, the analogue modem is either silent or sends TRN1a. The S that ends the DIL is looked for after either.

Phase 4¶

• In: CPt, CP and E.
• Out: Ri, TRN2d, MP, Ed and B1d go out through the encoder, first with the CPt constellations and then with the CP ones.

Data mode¶

Downstream is the encoder of 5.4/V.90: the modulus encoder, the mappers, the sign coding, and the spectral shaper with its look-ahead. The shaper makes the exact choice of 5.4.5.5, because the analogue modem checks TRN2d against it.

Upstream is the V.34 data mode receiver, with the trellis code and the precoding that our MP asks for.

The echo canceller¶

The echo canceller takes our PCM signal (converted to 9600 Hz) as its reference. It trains while Sd, TRN1d and Jd go out, because the analogue modem is meant to be silent then.

It stays off unless that training took at least 6 dB away (bypass in dp_v34.cpp). On a 4-wire network there is no echo, and a canceller trained on the remote's signal would only add noise.

With netsim's hybrid echo:

The precoding coefficients in MP¶

Our MP's precoding coefficients are found once per exchange, from the equaliser's error since the echo canceller took over. Before that point our own echo colours the error, and the predictor reads a gain of 3 to 6 dB into it.

Every MP of the exchange asks for the same coefficients, and the data mode receiver uses those. This matters because the HSF modem keeps what it was asked for first. When our receiver worked out its own coefficients at the start of the data mode (as it once did), its mse was 0.66 instead of 0.14.

Against the HSF modem with echo¶

The test hsf_v90_echo puts the HSF modem's hybrid echo 20 dB down and 50 ms late (NET_ECHO=-20:50):

• The canceller takes about 30 dB away (the test accepts 20 to 39 dB).
• The call connects at 52000/33600, sometimes after a retrain.
• With the echo 10 dB down it also connects, with a data mode mse of 0.18 against 0.14 without echo. The difference is the G.711 quantisation of the echo together with the upstream signal.
• With the echo after 20 ms instead of 50, netsim left a sixth of the echo out, and the upstream renegotiated down to 28800 or lower. The echo delay has to be longer than the modem sends ahead of what it receives; see NET_ECHO in testing.

The retrains are the HSF modem's own. After its S in phase 3 it goes silent without sending the S-bar that answers our Jd', then sends tone A. It takes no echo into itself; what changes is when we answer it:

• 20 dB down: our V.21 receiver gets the carrier some milliseconds later, with the echo of our ANSam on the line. It misses the first CM because it lacks all ten ones before the sync, so it sends JM 300 ms later. Accepting the sync after six ones would fix that. But the SM56 modem (whose first CM we would then catch too) chooses 45333 downstream instead of 50666 when JM comes one CM earlier, at every alignment tried. So the sync detector still wants all ten ones.
• 10 dB down: the HSF modem retrains even with JM on time. What we send differs from a call without echo only by a millisecond in phase 2. The call connects after one or two retrains.

Interop with analogue modems¶

The Smart Link library¶

The original library, as an analogue V.90 modem calling through the simulated network (ctest -R smartlink_v90), connects at 56000 bit/s downstream and 31200 upstream, in µ-law and A-law, on a line without noise.

Its phase 4 check works like this:

1. It compares the error of its equaliser output against the levels it expects for TRN2d's codes, over the first 7200 symbols of TRN2d.
2. In between, it replaces the levels it took from the DIL with those it measured in TRN2d.
3. It compares the error over the next 7200 symbols. If the second is more than about 1.5 times the first, it retrains (silence and tone A). The digital modem follows it back to phase 2 at the tones, as in a V.34 retrain.

On a line with no noise at all, the error is only a unit or two of its 16-bit scale. Its levels from the DIL are off by about that much (they are 4 times the G.711 values of the codes, give or take 2). So with µ-law the check fails every time: 1.96 against 1.23, and 1.78 on the second training, which it lets pass (it retrains at most once). The codes come through exactly as sent, so any digital modem would see the same.

With a noise floor of −81 dBFS (netsim's NET_FLOOR=3, still far below a real line's), the error is that of the noise. The ratio is then 1.2 to 1.3 at every alignment of the call in both laws, and the call connects at 54666/31200 without the retrain, 15 s sooner. The µ-law V.90 tests (smartlink_v90_ulaw, smartlink_v90_reneg, smartlink_v90_reneg_silence) use that floor.

The original library's own debug output helped here (its edprintf lines are hidden behind a key). It reports the levels it measures in the DIL, the digital pad it concludes and why it retrains.

Other analogue modems¶

The Agere modem as the analogue modem (ctest -R agere_v90) connects at 46666 bit/s downstream. It trains on TRN1d for 0.6 s and then decides on it for another 0.9 s before it looks for Jd. That is why TRN1d lasts 1 s; it used to be 300 ms, which was enough only for the original library.

Over SIP¶

Through Asterisk and DModem (frontend/asterisk/test/sip.sh AT+MS=90, with CODEC=alaw for A-law), the original library connected at 56000/31200 in both laws. That was while dmodemd (then slmodemd) still ran the original library at 9600 Hz, and its SIP side did the 8000/9600 Hz conversion itself. (pjmedia's resampler cuts below 4000 Hz, which takes away the part of the PCM signal that the analogue modem recovers its timing from.) dmodemd now runs at the call's 8000 Hz.

Rate renegotiation¶

Rate renegotiation (9.6/V.90) changes the rates in the middle of a call without a full retrain.

The analogue modem's¶

The analogue modem sends S, S-bar, SCR and CP with the downstream rate it wants. We answer its S-bar with:

1. Rd: + + + − − − of each interval's largest data mode Ucode, for 384 symbols, on a data frame boundary; then Rd-bar (the opposite) for 24 symbols.
2. TRN2d.
3. MP, with the upstream rate our mse allows; then MP'.
4. Ed and B1d.

The encoder uses CPt's constellations until B1d.

Ours¶

Ours start from the upstream mse, as in V.34, with Rd first. DSP_V34_RENEG=<ms>:<N>[a|c] starts one in a test: <ms> into the data mode, to N × 2400 bit/s.

The DTE can also ask for one with ATO3 (and for a retrain with ATO1); see open_v90_dte_reneg in dsp/test/interop_test.cpp.

The silent period (CPs)¶

A CP with bit 30 set (CPs) asks for a silent period, so that the analogue modem can retrain its echo canceller. We then send:

1. MP' until CPs' arrives, then Ed;
2. Ucode 0, keeping the data frames, until a CP with bit 30 clear arrives;
3. Rt (CPt's largest Ucodes), Rt-bar and MP.

If no E arrives within 5 s of Rd (plus the round trips), we retrain.

Tests¶

The original library renegotiates on its own when the downstream gets worse, always with the silent period. With the noise still there it now and then picks 56000 again, and its data then has errors. So the test stops netsim's noise during our silence (NET_NOISE_QUIET).

The Agere modem renegotiates on its own as well, without the silent period, a step down at a time as the noise rises (45333, 46666, 44000 in one call). But it only does so after 40 s or so, if at all, so there is no test for it.