V.61 (ASVD): technical details¶

The plain-English overview is v61.md. The references are to ITU-T V.61 (08/96) and V.8 bis (11/2000).

The parts¶

V.8 bis¶

With +ASVD=1 the CL's SVD mode has V.61, V.42 and V.14 (8.3.2/V.8 bis). The caller's MS chooses DSVD when both have what V.70 needs (see v70-details.md), else ASVD when both have V.61: the SVD mode with V.61 and the V.42 and V.14 bits both have, ACK(1) asked for, no V.8 (the I field's V.8 bit clear: V.61 has its own start-up). The answerer takes an MS with V.61 and not V.70 only if it offered ASVD, with ACK(1); else NAK(3) and V.8 without V.8 bis.

Then both ends ask the core for DP_V61 (MDMPRM_DP_REQUESTED) in place of V.8: the answerer once its ACK(1) is out, the caller at the ACK(1), or at 100 ms of the answerer's 2100 Hz if the ACK(1) was lost. The V.61 answerer then starts with its own ANS.

The line signal¶

Both have 2400 data symbols a second, so the rates are the same at both: 2 to 6 bits a symbol for 4800 to 14400 bit/s. The transmit pulse is a square root raised cosine of roll off 0.12 (Figure 7), over 16 symbols each side: at 6, its truncation left inter-symbol interference at about -30 dB, which 64 points do not survive.

Data alone is uncoded QAM, the 4, 8, 16, 32 and 64 points of Figures 1 to 5 (read from the PDF's figures with pdftotext -bbox: they nest, and each is unchanged by a turn of 90 degrees), with Table 1's differential quadrant code on the first two bits of each symbol.

With speech (QADM, 5.5) a frame has the 4 points of Figure 6, at 2.5 dB under the data-alone 4 points' level, 2 bits a symbol (4800 bit/s), and the speech added around them:

1. The speech moved down by the audio carrier, low pass filtered and taken at each symbol (V61AudioTx: a Kaiser windowed sinc, 88 taps each side at 8000 Hz, evaluated at the symbol instants directly, as the note to 5.5.1.2 allows).
2. The frame's gain (5.5.2.3): 2^(1/4) steps, 0 to 31 for the smaller of the inphase and quadrature (the base) and up to 3 more for the other, chosen from the frame's peaks ahead of it.
3. Every other symbol inverted (5.5.2.4).
4. Each part scaled asymmetrically (5.5.2.5): a negative half that reaches 0.7 of the way to the quadrant's edge (0.7 * sqrt 5 times the points' scale), a positive half twice that.
5. Turned by the data point's angle less 45 degrees (Table 5), so the negative half points toward the quadrant's edges.

The receiver decides the quadrant (the data point nearest in angle) and undoes 5 to 2 on the rest: that is the speech. V61AudioRx interpolates it back to 8000 Hz and moves it up by the audio carrier. Both ends count the audio carrier's phase from the first symbol after E.

What D-Modem chooses¶

Control segments¶

Table 6's bits, one dibit a control symbol: symbol k carries bit 2k+2 first, then 2k+1. In data alone the third bit of each control symbol is a copy of D, the rest 0 (5.6.2). The receiver checks the inverted copies (D and D-bar, L and L-bar, G4 to G2 and their inverses, the fixed bits) and decides D by the majority of D, not D-bar and the copies, a tie being data alone (Note 1 to Table 6). After a data-alone frame the gain is 0 and the lock bit counts as set.

Start-up¶

R1, R2 and R3 are rate words 1, 2 and 3 (Tables 10 to 12), sent one after the other and repeated: 16 bits each, most significant (B0) first, header 0000 for R and 1111 for E, sync bits 0111, rate word 1's ID bits B4 and B8 set, rate word 2's B4, rate word 3's B8. Rate word 1 has the rates, rate word 2 bit B5 (A+D) where the sender has a voice path, rate word 3 bit B14 (silence switching). The caller scrambles with GPC (tap 18), the answerer with GPA (tap 5).

Each receiver offers the rates its training measured:

The echo canceller (256 taps) trains on our own TRN. Its step is 0.25, then 0.05 halfway through TRN and 0.02 at three quarters: with a far echo at -15 dB it takes 28 dB away.

The equalizer (48 taps at T/2) adapts on decisions in TRN (a step of 0.04 while its error is over 0.01, then 0.02), on known symbols in B1 (0.04: the far end's scrambled ones, predicted from our descrambler) and on decisions in data alone (0.02). With speech it adapts slowly (0.002) while the lock bit is clear and not at all while it is set; the carrier loop follows throughout. With a smaller step in data (0.004) the error stayed at -24 dB for seconds after the start, too high for 64 points.

Retrain¶

On the far end's G or F for 128T in data, or when the receiver has lost its equalization while the signal is still there (its error 4 dB short of the rate's SNR for 2 s, 7 dB at 4800 bit/s, or the control segments failing their checks for 1 s), the caller sends G and the answerer F, and the start-up runs again from step 4.

Tests¶

Not done¶

• Rate renegotiation (clause 11): a retrain instead.
• Adaptive pre-emphasis (5.5.1.2): the flat filter only.
• The auxiliary control channel (5.6.3).
• Changing to ASVD during a telephone call (V.8 bis in the middle of a conversation).
• Tried only between two D-Modems.