GopherTrunk — Decoders that still need live RF captures

A walk through every decoder in the tree, flagging which ones are blocked on a real over-the-air capture before they can be called fully validated, and exactly what to record. Everything below decodes synthetic IQ today — the gap is confirming the on-air constants (bit order, deviation, interleave schedule, FEC margins) against a real signal.

⚠️ Read this first: IQ vs. audio

GopherTrunk’s pipelines start at complex IQ baseband. For anything that carries information in phase or 4-level amplitude, an MP3/WAV of the demodulated audio is useless — the constellation is already gone.

  • Must be IQ (.cfile / .bin / .iq, or a stereo IQ .wav with I=left / Q=right): TETRA, NXDN, YSF, DMR, M17, P25.
  • Audio is OK only where the data rides audio-band FFSK tones: MPT 1327 (and as a fallback for POCSAG/FLEX/DSC, though IQ is preferred).
  • .cfile = GNU Radio interleaved float32 IQ (8 bytes/sample). .bin = complex int16 (4 bytes/sample). Neither embeds the sample rate or center frequency — always drop a metadata.json sidecar stating sample_rate_hz + center_freq_hz.

Drop captures in samples/<protocol>/ with a metadata.json. Skip-gated integration tests pick them up automatically.

🔴 Tier 1 — Blocking: decoder unproven on air, harness ready

These have a finished decode chain and a waiting test harness, but zero real-air confirmation. A single labeled capture closes each one.

NXDN (NXDN-TS-1-A control channel)

  • What to capture: Outbound RCCH control-channel carrier.
  • Format: complex float32 .cfile (or int16 .bin).
  • Sample rate: ≥ 48 kHz (48 kHz nominal).
  • Modulation: 4-level FSK @ 4800 sym/s, 6.25/12.5 kHz channel.
  • Duration / size: ≥ 5 s (~2 MB at 48 kHz f32); 10 s is better.
  • Specifics: spec peak deviation is 1800 Hz — if the dibit distribution comes out lopsided, sweep nxdn_deviation_hz (non-spec rigs land ~2200–2700 Hz). Note: an NXDN-BFSK (2-level, 4800 bps) capture will not decode — the receiver is 4-FSK only.
  • Cross-check: MMDVMHost log + DSDcc 1.9.5.
  • Pass bar: ≥ 80% CAC bursts CRC-OK, SystemID/SiteID/RAN byte-match, CC lock < 3 s.
  • What to capture: BS downlink carrier (SCH/HD, SCH/F, BSCH frames).
  • Format: complex float32 .cfile (or int16 .bin).
  • Sample rate: ≥ 72 kHz strongly recommended (≥ 4 samples/symbol at 18 ksym/s for reliable Gardner lock — a 48 kHz capture is too low and will not lock).
  • Modulation: π/4-DQPSK @ 18 ksym/s, 25 kHz channel.
  • Duration / size: ≥ 30 s (~17 MB at 72 kHz f32) — include an idle window for noise-floor profiling.
  • Specifics: cleartext frames only — TEA1–4 encrypted traffic can’t be used (key recovery is out of scope). Co-channel / adjacent- channel interference is welcome; that’s exactly the Viterbi margin the capture is meant to profile.
  • Cross-check: telive 1.5 / osmo-tetra.
  • Pass bar: CC lock < 5 s, ≥ 90% frame recovery, Viterbi correction depth p95 ≤ 8 / p99 ≤ 12 bit-errors per 116-bit block.
  • Verification progress (live 468.5 MHz / 1 Msps captures): two blockers found and one fixed. (1) The captured front-end was spectrum-inverted — set iq_invert: true (or replay with -conjugate); with that the normal-burst demod is clean (NTS1 at Hamming distance 0). (2) A real-air decoder bug: the BSCH colour-code recovery only correlated the rotation-0 orientation of the synchronisation training sequence, but π/4-DQPSK’s residual-CFO term rotates the whole dibit stream — now fixed to try all four rotations (internal/radio/tetra/process.go, guarded by TestRecoverColourCodeUnderRotation + TestProcessLearnsColourCodeFromSBBurstUnderRotation). Still open: the supplied clips are only ≈3.3 s each (too few frame-18 SB slots) and their SB region does not demodulate cleanly, so no cc.locked fires yet. Closing this needs a ≥30 s cleartext capture spanning several clean synchronisation bursts. See samples/tetra/README.md.

YSF (Yaesu System Fusion, DN mode)

  • What to capture: A YSF carrier (Pi-Star/reflector or a keyed HT).
  • Format: complex float32 .cfile / int16 .bin (or stereo IQ WAV).
  • Sample rate: ≥ 48 kHz nominal.
  • Modulation: C4FSK @ 4800 sym/s, ±2700 Hz peak deviation, 12.5 kHz.
  • Duration / size: ≥ 10 s (~4 MB at 48 kHz f32) — enough FICH cycles.
  • Specifics: confirms the FICH interleave/puncture schedule ({0,1,102,103} + column-major 10×10, per MMDVMHost). If CRC fails there’s a documented two-line swap to the DSDcc alternate schedule — if so, publish the K=5 generator pair you landed on.
  • Cross-check: DSDcc 1.9.5 in YSF mode / Pi-Star FICH log.
  • Pass bar: 100% FICH CRC at clean SNR, ≤ 4 trellis bit-errors per 100-bit block at ≥ 12 dB.

DMR 2-slot interleaved voice (opt-in dmr_interleaved_voice)

  • What to capture: A live DMR carrier with both timeslots active (Tier III trunked voice, or a busy Tier II repeater).
  • Format: GNU Radio interleaved float32 .cfile.
  • Sample rate: state it in the env/sidecar (48 kHz works with the harness).
  • Modulation: C4FM @ 4800 sym/s.
  • Duration / size: long enough for a full call on the followed TG (≥ 10–15 s, ~4–6 MB at 48 kHz).
  • Specifics: this is the gate before the interleaved decoder can replace the single-slot decoder as the default. It confirms two unknowns: (1) the same-slot dibit cadence on live BS air — whether a CACH makes the stride 288 vs. 264 dibits; (2) the ETSI embedded-LC de-interleave order, EMB QR(16,7) FEC, and 5-bit CRC polynomial. Run via GOPHERTRUNK_DMR_2SLOT_CFILE / _RATE_HZ / _TG.
  • Pass bar: the followed talkgroup’s embedded LC is recovered and its (and only its) audio is routed to the sidecar.

DMR Full-LC FEC on air (issue #527 follow-up) — ✅ CONFIRMED

  • Status: Closed by live capture. A set of 441 MHz / 2 MS/s GNU Radio f32 captures of a DMR Tier III trunked system was channelised and replayed through the production receiver. The receiver’s real-air dibit recovery and end-to-end bit ordering through BPTC(196,96) → RS(12,9) → Full LC parse — the exact stack the Voice LC Header uses and the only thing the field report left unproven — decode cleanly on air.
  • What settled it: the Voice LC Header and the Terminator-with-LC carry the same Full LC PDU through the same BPTC + RS(12,9) FEC (only the RS seed differs, 0x96 vs 0x99, both pinned by TestRS129MatchesIndependentReferenceEncoder). On-air Terminator-with-LC bursts recover a stable RS-validated FLC (TG 24 / source 4209000), and the control channel’s CSBKs pass BPTC + CRC ~97–98%, with the Tier III ControlChannel locking on the Aloha beacon.
  • Regression guards: TestReplayDMRVoiceLCFECDecodesRealAir + TestReplayDMRTier3ControlDecodesRealAir in cmd/gophertrunk/dmr_realcapture_test.go (fixtures cmd/gophertrunk/testdata/dmr-{voice-term,t3-cc}.cfile).
  • Root cause of the original field failure: the unit-energy RRC matched filter overdrove the fixed 4-level slicer, collapsing inner symbols onto the outer rails so every BPTC payload came back uncorrectable while the more forgiving sync + slot-type FEC still passed — exactly the reported symptom. The symbol-AGC calibration in internal/radio/dmr/receiver (issue #275) fixed it; these captures confirm the fix on real bits.

DMR Tier II Voice LC Header (conventional repeater) — 🟡 optional

  • What to capture: A live conventional DMR repeater carrying voice — key-up to un-key, so the capture spans a Voice LC Header and the Terminator-with-LC that bracket a transmission.
  • Format: complex int16 .bin or GNU Radio float32 .cfile under samples/dmr-tier2/. Not demodulated audio.
  • Sample rate: ≥ 48 kHz. C4FM @ 4800 sym/s, ~1944 Hz deviation, 12.5 kHz.
  • Duration / size: ≥ 5 s (~1 MB int16) — one call is enough.
  • No longer blocking: the underlying real-air BPTC/RS Full-LC concern is resolved above. A conventional capture now only adds direct coverage of the Tier II call-setup path (Voice LC Header → grant → Terminator) rather than the trunked Terminator-with-LC that shares its FEC.
  • Cross-check: MMDVMHost log / DSD-FME / radio display.
  • Pass bar: the captured Voice LC Header decodes to the expected talkgroup + source ID + color code, with no decode.error stream on a clean-SNR call.

🟡 Tier 2 — Calibration: works on synthetic, needs real to lock constants

Decoders that run end-to-end against a synthetic encoder; a real capture calibrates timing/deviation/level constants or confirms bit order.

IMBE vocoder (P25 Phase 1 voice level calibration)

  • What to capture: A 5 s+ P25 Phase 1 voice call, recorded with recordings.write_raw: true.
  • Deliverables (two files, same call):
    • p25-p1-voice.raw — raw IMBE frames, 11 bytes/frame, MSB-first packed, no header (the daemon’s .raw sidecar).
    • p25-p1-voice-dsdfme.wav — the same .raw run through DSD-FME / OP25, 8 kHz / 16-bit / mono PCM.
  • Size: tiny (raw frames + a short WAV).
  • Purpose: lets calibrate.Compare quantify the AGC/level offset.
  • Pass bar: RMS ratio < 3 dB, peak cross-correlation > 0.85.

AMBE+2 vocoder (DMR / NXDN / dPMR / D-STAR voice level calibration)

  • What to capture: A voice call on any AMBE+2 protocol, write_raw: true.
  • Deliverables (two files, same call):
    • dmr-voice.raw — raw AMBE+2 frames, 7 bytes/frame, MSB-first.
    • dmr-voice-dsdfme.wav — same .raw via DSD-FME / OP25, 8 kHz/16-bit/mono.
  • Specifics: Knox / call-alert tones (b₁ ∈ [144,163]) are vendor- specific and decode as silence unless registered — fine, just note it.
  • Pass bar: RMS ratio < 3 dB, peak cross-correlation > 0.85.

POCSAG paging

  • What to capture: A POCSAG paging channel.
  • Format: complex float32 .cfile under samples/pocsag/.
  • Sample rate: ≥ 48 kHz.
  • Modulation: 2-FSK / FFSK on NBFM; 512 / 1200 / 2400 bps.
  • Duration: ≥ 30 s to catch several batches (~12 MB at 48 kHz f32).
  • Purpose: tune the integrator/slicer timing against real bits (the running-mean slicer likely becomes a proper Mueller-Müller + matched filter once there’s signal to calibrate against).

FLEX paging

  • What to capture: A FLEX paging channel.
  • Format: complex float32 .cfile under samples/flex/.
  • Sample rate: ≥ 48 kHz.
  • Modulation: 1600 bps / 2-level mode (sync 0xA6C6AAAA, mode 0x870C).
  • Duration: ≥ 30 s.
  • Purpose: confirm sync framing, BCH(31,21) bit order, and capcode mapping on air (3200/6400 bps and 4-level modes are out of scope).

DSC (marine VHF Ch-70 distress)

  • What to capture: A real ITU-R M.493 DSC burst on channel 70.
  • Format: IQ .cfile (FFSK rides the audio band, but IQ preferred).
  • Modulation: 1200 Bd FFSK, 1300/2100 Hz tones on NBFM.
  • Purpose: confirm on-wire bit order, tone sense, and DX/RX time- diversity offset against a captured signal.

APRS / AX.25 packet

  • What to capture: A real APRS channel (144.390 in NA).
  • Format: .wav or .cfile under samples/aprs/.
  • Modulation: Bell-202 AFSK, 1200 bps.
  • Duration: long enough for several beacons (~30–60 s).
  • Purpose: replay through internal/sdr/baseband to assert the full AFSK → HDLC → AX.25 → APRS chain (only the synthetic test exists today).

AIS (marine vessel positions, 161.975 / 162.025 MHz)

  • What to capture: Marine VHF 87B/88B.
  • Format: IQ recording under samples/ais/.
  • Modulation: 9600 Bd GMSK (BT ≈ 0.4).
  • Purpose: exercise the GMSK frontend on air (only the bit-stream synthetic is tested today). Single-slot messages only for now.
  • What to capture: An M17 transmission.
  • Format: IQ .cfile / stereo IQ WAV.
  • Modulation: 4FSK / C4FM @ 4800 sym/s.
  • Purpose: confirm the Golay matrix, C4FM slicer deviation, and symbol-timing constants (validated against a synthetic encoder so far). Codec2 voice is a separate future milestone, not a capture gap.

🟢 Tier 3 — Optional: pipeline closed, captures only add robustness

Already validated end-to-end (synthetic fixture passes through the production pipeline). Real captures are welcome for burst-error / false-positive coverage but not blocking.

MPT 1327 (control channel)

  • 48 kHz IQ .cfile or 8 kHz demod-audio .wav (FFSK is audio-band, so audio works here). FFSK 1200 baud, 1200/1800 Hz tones, NBFM. ≥ 60 s (~23 MB IQ / ~1 MB audio) to characterise the noisy-CWSC distribution. Closes empirical false-positive rate + per-vendor (Tait/Motorola/Simoco) sync bit-error patterns. Already decodes real audio end-to-end today.

Not capture-blocked (listed so the question is fully answered)

  • ADS-B — validated against the canonical dump1090 / mode-s.org reference vectors; no live capture outstanding.
  • MDC1200 — over-the-air FEC redundancy isn’t exploited yet, but that’s a decode-improvement follow-up, not a capture gap.
  • P25 Phase 1 / Phase 2 — full TIA-102 chains ship and decode to audio; Phase 2 inner FEC (trellis / RS / PN44) is closed. The Phase 1 C4FM control-channel demod is now confirmed on real air: a live UHF capture (449.875 MHz, NAC 0x2C1) was channelised to 48 kHz and run through the samples/p25/ demod-quality harness (TestReplayP25RealCaptureMetrics), which grades it at EVM ≈ 12.7%, SNR ≈ 14.5 dB, NID 31/0, TSBK 36/0. The capture binary is not committed (per samples/.gitignore); the committed record is the samples/p25/p25-450875-cc.metadata.json sidecar, and the capture is reproducible locally via TestGenerateP25Fixture. (The Tier 2 IMBE voice-level calibration above is a separate gap — it needs raw IMBE frames + a DSD-FME reference WAV, which a control-channel IQ capture does not provide.)
  • EDACS, LTR, Motorola Type II, dPMR control — control chains ship; FEC is on by default with no outstanding capture.

Separate from captures: digital-voice composer chains

NXDN, dPMR, TETRA, YSF, D-STAR voice (plus EDACS ProVoice) are followed and logged but not yet turned into PCM — that’s an implementation gap (vocoder/composer wiring), not something a capture unblocks.

TL;DR priority order for anyone with an antenna

  1. NXDN outbound control (≥ 5 s IQ, 48 kHz) — harness fully ready.
  2. TETRA TMO downlink (≥ 30 s IQ, ≥ 72 kHz, cleartext).
  3. YSF DN mode (≥ 10 s IQ, 48 kHz).
  4. DMR 2-slot both-timeslots-active call (≥ 10 s .cfile).
  5. IMBE + AMBE+2 voice .raw + DSD-FME WAV pairs (level calibration).
  6. POCSAG / FLEX / DSC / APRS / AIS / M17 real-fixture confirmation.

Every capture needs a metadata.json sidecar with sample_rate_hz, center_freq_hz, and the expected decode (SystemID / talkgroup / callsign / message text) so the harness can grade it, not just smoke-test it.