pico-serial WB — wideband bare-pair stick (draft 0.1)¶

A sibling of the phone-line stick (SPEC-L.md) for bare copper pairs: dry pairs, alarm and leased circuits, a pair of your own in a building or between buildings. Two WB sticks talk to each other point to point, sampling the pair at 96 kHz instead of a phone line's 8–9.6 kHz, so they use far more of the pair's bandwidth than any voice-band modem can.

Same as the Line stick: the stick is only the line interface (a codec, a transformer, protection) with an RP2040 streaming samples over USB, and the modem is D-Modem on the host (Linux dmodemd, Android's dmodem app), here a new wideband mode of our own (§6). Both ends are ours, so the modulation needn't follow any standard.

Not covered: telephone lines (use the Line stick: a phone line's exchange passes only 300–3400 Hz, whatever the modem), ADSL and SHDSL (megahertz bands, standard line codes, a carrier's equipment at the far end: see §9).

1. Requirements¶

2. Block diagram¶

 USB-C ── ESD ── RP2040 ── I2S (PIO) ──► TLV320AIC3204 ── line out ─► R ─┐
  (UFP)          │  MCLK 12 MHz (GPOUT) ►  codec, PLL      ◄─ line in ◄─ hybrid
  LDO 3.3 V      │  I2C ───────────────►   96/192 kHz                     │
  flash, xtal    └── LEDs (LINK, DATA)                         ═══ transformer ═══ DC block ─ fuse ─ RJ11
                                                                isolation barrier   caps    sidactor

3. The line interface¶

• Codec: TI TLV320AIC3204 (LCSC C24109): stereo ADC and DAC to 192 kHz (but run at 96 kHz, B4), its own PLL (so the RP2040's 12 MHz crystal clock is its MCLK and it is the I2S master, as the Si3056 is the serial port's master on the Line stick), a headphone driver strong enough for a line transformer, differential inputs, I2C control, one 3.3 V supply (internal LDO).
• Transformer: a wideband telecom transformer (the ISDN-U / HDSL / SHDSL kind: a few kHz to hundreds of kHz) is the isolation barrier, as the Si3056's capacitors are on the Line stick. Parts: see §8.
• DC block: capacitors in series with the line side (≥ 250 V), so a live phone line's −48 V doesn't saturate the transformer.
• Line drive: the codec's headphone outputs, differential (HPL = s, HPR = −s), through 60.4 Ω in each leg: 120 Ω, about the pair's impedance; about +3 dBm on the pair at full scale (TI's guide supports a differential load only with the drivers on AVDD, 1.0 Vrms between HPL and HPR). No op-amp.
• Hybrid: frequency division keeps the far end's band apart from our own, but our own transmit level is far above what we receive. A replica arm at ten times the impedance (604 Ω each leg, and 1.21 kΩ in series with 10 nF for the cable) makes a copy of what we send; the codec's input mixer subtracts it (P = IN1L + IN1R, M = IN2R + IN3R, 10 kΩ each), taking 10–20 dB of our own signal out before the ADC. The modem's filters take the rest.
• Protection: the Line stick's sidactor and telecom fuse on the line side, clamps on the codec side.

The netlist variant is --variant wb; the shell is the Line stick's (stick.py), with the LEDs relabelled.

4. Clocks and rates¶

The codec's PLL makes its 96 kHz from the RP2040's 12 MHz crystal clock (GPOUT, no fractional divider): P = 1, R = 1, J.D = 8.1920 gives 98.304 MHz; NDAC = NADC = 4, MDAC = MADC = 4, DOSR = AOSR = 64 (blocks PRB_P7 and PRB_R7); BCLK = DAC_CLK / 4 = 6.144 MHz, 64 bits a frame. For 192 kHz (tests): DOSR = AOSR = 32, PRB_P17 and PRB_R13, BCLK / 2. The codec is the I2S master: the RP2040's PIO takes its bit clock and word clock as the Line stick's PIO takes the Si3056's. So the sample clock is the stick's crystal, and the USB side follows it.

Two sticks never share a clock: their crystals differ by up to about ±50 ppm, and the modem's receiver tracks the difference (§6).

5. USB¶

As the Line stick (PROTOCOL.md, "Line stick"): one vendor interface, isochronous IN and OUT, 1 ms packets sized by the codec's clock, an interrupt endpoint for events. At 96 kHz S16 that is 192 bytes a millisecond each way (384 at 192 kHz), well within Full Speed's 1023-byte isochronous packets. A distinct PID (pid.codes test PID) and a version bit say which stick it is. Vendor requests: info, rate, codec register access, the line's level and a loopback test; no hook, ring or caller ID.

6. The modem: a wideband mode for D-Modem¶

A new data pump in D-Modem (dsp/), with D-Modem's error correction and compression above it (V.42 LAPM, V.44) as for every other mode.

Multicarrier (DMT/OFDM), frequency-division duplex. - At 96 kHz: a 512-point FFT gives 256 tones 187.5 Hz apart; a cyclic prefix of 40 samples; about 174 symbols a second. Tones about 21 to 192 (4–36 kHz) are usable (B4). - The two directions on separate tones: the end that dialled sends on the lower band, the end that answered on the upper, with a guard band between. The split is set at training from both directions' measured capacity, so the two directions get about the same rate. - Per-tone bit loading: each receiver measures every tone's signal-to-noise ratio during training and tells the sender how many bits each tone carries (0 to 12) and at what gain, as ADSL does. Long pairs lose their top tones; short ones use them all. - Pilot tone and clock tracking: a pilot in each direction lets the receiver follow the two crystals' difference (rotating each tone's phase and slipping a sample now and then). - Coding: Reed-Solomon (D-Modem's dsp/rs.cpp) with interleaving; trellis coding later if it pays. - Start-up: the caller sends a probe tone; the answerer (auto-answer on a leased line, AT&L1, or ATA) replies; then the training sequence (sync symbols, the SNR measurement, the bit tables exchanged on a robust 2 bits-per-tone channel), then data. Retrain on loss of sync, and adapt the bit tables in a call (bit swapping) as the pair's noise changes.

Rates measured on a simulated pair (D-Modem vwb_test: 24 AWG cable loss, −15 dB hybrid echo, crystals 100 ppm apart, no bit errors), each way at 96 kHz:

Through V.42 both ways at once: 137.6 kbit/s of data each way on 0.5 km. Not yet tried on a real pair; the stick's transmit level (§3, about +3 dBm) and its noise will set the real figures.

Leased-line behaviour (Hayes &L1): no dial tone, ringing or hook. ATD (any number) starts as the caller; ATA, or ATS0=1, waits for a caller's probe. Carrier lost: NO CARRIER, and with &L1 the modem tries again by itself.

Test plan: both pumps against each other over a simulated pair: frequency-dependent loss (a cable model), noise (white and narrowband), the hybrid's leakage, crystals 100 ppm apart; then two sticks on a real pair (a drum of telephone cable).

7. Host software¶

• dmodemd --usb finds a WB stick beside the SmartUSB56 and the Line stick; -r 96000 (the default for it).
• The core runs at the line's rate: D-Modem's pumps already take their sample rate from the core (modem_dp srate); the line drivers' and options' 8000/9600-only checks widen for this device.
• dmodem-app: the same, through Android's USB host API and libusb.

8. Parts¶

Checked against JLCPCB/LCSC stock on 2026-10-27.

Plugged into a live phone line by mistake: no DC path, so the exchange sees it on-hook; ringing drives about 6 mA through it.

To check on the bench: the transformer's datasheet gives no bandwidth curve or saturation rating (sweep one from 1 to 200 kHz), and the hybrid values are a starting point from typical cable data.

9. Not this stick: ADSL and SHDSL¶

ADSL uses 25 kHz–1.1 MHz (ADSL2+ 2.2 MHz): about 4.4 million samples a second each way, past the RP2040's Full Speed USB, and it would need an ADSL analogue front end and, to reach an ISP, the whole G.992 standard (DMT, its training, ATM) against a DSLAM. SHDSL likewise. A study, on a different chip, if wanted later.

10. Next steps¶

1. Parts (§8), then the netlist, schematic and PCB (--variant wb).
2. Firmware: the codec over I2C, I2S in PIO, the Line stick's USB at 96/192 kHz.
3. The modem in D-Modem, developed against the simulated pair first.
4. Two sticks on a real pair.