Also known as: TETRA bring-up facts, TETRA sync gotchas
TETRA lock facts are the small set of protocol and DSP truths that separate a TETRA demodulator that locks real air from one that only passes its own tests. Each was learned during GopherTrunk’s TETRA bring-up, and several contradict the “obvious” reading of the spec.
Training sequences are counted in bits, not dibits
ETSI EN 300 392-2 §9.4.4.3 defines the training sequences as NTS1/NTS2 = 22 bits, ETS = 30 bits, STS = 38 bits. The trap is unit confusion: “38” is bits (19 dibits), not 38 dibits. GopherTrunk’s first implementation carried placeholder constants that declared a 64-bit variable to hold “38 dibits” (76 bits) — six dibits were silently zero-filled, and the correlator could never fire on real signal (#553).
The decisive proof that a sync layer works, independent of anything downstream: training correlation hits at a modal spacing of exactly 1020 dibits — one TETRA frame (4 slots × 255 dibits). Hits at exact frame cadence validate sync even while every CRC still fails.
π/4-DQPSK uses the Gray mapping
On air, a π/4-DQPSK symbol pair (B1, B2) maps to a dibit via the
Gray convention (B1<<1) | (B1^B2):
| Bits | Dibit |
|---|---|
| 00 | 0 |
| 01 | 1 |
| 11 | 2 |
| 10 | 3 |
Using the linear convention (correct for the C4FM family) flips B2 whenever B1 = 1. Sync still correlates, but descrambling, deinterleaving, and CRC all see corrupted bits — a “locks but nothing decodes” state (#553).
AFC: the 4th-power estimator aliases at ±f_sym/8
Raising the per-symbol differential phase to the 4th power collapses π/4-DQPSK data to a constant — but because 4·ω wraps at 2π, the offset is only recoverable modulo π/2 per symbol. At TETRA’s 18,000 sym/s that is ±f_sym/8 = ±2250 Hz. Larger offsets alias: the estimator returns a plausible-looking value that is wrong by a multiple of f_sym/4 = 4500 Hz (a 2500 Hz offset reads as ≈ −2000 Hz), and lock silently fails (#940).
The fix is a coarse acquisition stage — the angle of the block’s lag-1 autocorrelation, which has no ±π/4 wrap — that picks the correct 4500 Hz alias bucket before the 4th-power stage refines. Two subtleties:
- The coarse stage must read pre-matched-filter samples. The matched filter is centred at 0 Hz, so a spectral estimate taken after it is pulled toward zero (measured at roughly half the true offset).
- Total range is bounded near ±6 kHz by the 15 kHz channel-filter cutoff.
Also note ordering: AFC has to run before timing recovery — an off-centre channel biases every dibit and the training sequence never correlates in the first place (#553).
Colour code 0 is not “no scrambling”
TETRA scrambling is non-identity at colour code 0: the
colour-0 scrambler seeds its LFSR to 0xC0000000 (EN 300 392-2 §8.2.5.2). Two consequences
that each cost real debugging time:
- Blind identification must recover the colour code from the BSCH. The BSCH is always transmitted colour-0-scrambled and is therefore always decodable; every other channel needs the site’s real colour code. GopherTrunk’s hunt scored TETRA candidates with the configured colour (0 during blind identify), so every SCH/HD CRC failed and the confidence score forfeited its entire FEC term — recover the colour from the BSCH first (#648).
- Never skip descrambling “because the colour is 0.” The DMO (direct mode) voice path
inherited an
if colour != 0shortcut from the trunked-mode code — safe there only because a trunked extended colour code is never 0. On clear colour-0 DMO traffic it left bursts scrambled going into the FEC, producing a uniform chance-floor CRC rate that was misread as encryption (#1003).
A related DMO fact: the DM colour code is not recoverable from the SCH/S (which is always colour-0-scrambled and so always reads 0). When signalling and radio programming disagree with reality, GopherTrunk recovers the colour empirically — sweep candidate colours and keep the one that maximizes CRC-valid speech frames; the correct colour wins by a wide margin (#1003).
Symptom table
| Symptom | Looks like | Actually | Fix / check |
|---|---|---|---|
| Training sequence never correlates on real air | Wrong sample rate, inverted spectrum | Placeholder/miscounted sync constants, or no AFC before timing | Bit lengths 22/30/38; look for hits at 1020-dibit cadence (#553) |
| Sync locks, every CRC fails | Scrambler or colour bug | Linear instead of Gray dibit mapping | Apply the Gray table (#553) |
| AFC reports a small, stable offset but never locks | AFC converged, signal weak | 4th-power alias — residual is a multiple of 4500 Hz | Coarse lag-1 autocorrelation stage, pre-matched-filter (#940) |
| Colour 0, so descrambling skipped | Harmless shortcut | Colour-0 scrambling is non-identity (seed 0xC0000000) |
Descramble unconditionally (#1003) |
| Known-clear traffic decodes at the CRC chance floor | Encryption | Wrong colour code for the descramble | Recover colour from the BSCH, or sweep for max CRC yield (#648, #1003) |
The DMR sync layer has an analogous blind spot — its sync words are closed under polarity flip, so sync alone can’t reveal spectrum inversion either; see DMR grant-field gotchas and the wider catalog in signal signatures.
Provenance
- #553 — TETRA fails to lock: placeholder training sequences, the bit-vs-dibit trap, the Gray mapping, and the frame-cadence proof.
- #940 — AFC coarse-acquisition range: the ±2250 Hz 4th-power alias and the pre-matched-filter rule.
- #648 — hunt misclassification: blind identify at colour 0 forfeits all FEC confidence; recover the colour from the BSCH.
- #1003 — DMO clear voice misread as encrypted: the colour-0 descramble skip and empirical colour recovery.