Status & known gaps

This page tracks what’s shipping end-to-end versus where the engine follows a call but doesn’t yet turn it into audio (or doesn’t yet correct every on-air FEC layer). The high-level summary lives in the README; this page is the long-form reference.

What ships today

Once a grant event lands on the bus, the engine + recorder pipeline runs end-to-end: voice device is allocated, the composer pulls IQ → PCM, the recorder writes a WAV (digital-voice protocols decode through the right vocoder via voice.DefaultVocoderForProtocol), the call is logged to SQLite, and the API + TUI surfaces all light up. Pure-Go IMBE / AMBE+2 and the clean-room TETRA ACELP vocoder produce intelligible audio. The CC Hunter supervisor and the conventional FM scanner are constructed by cmd/gophertrunk and expose their state through /api/v1/scanner and the TUI cockpit panel.

Every trunked control modulation in the Features table has an end-to-end IQ → CC chain shipping. The ccdecoder connector covers all 10 trunked protocols (P25 Phase 1, P25 Phase 2, DMR Tier III, NXDN, dPMR Mode 3, EDACS, Motorola Type II, LTR, MPT 1327, TETRA TMO) plus DMR Tier II conventional and YSF / D-STAR on the amateur side.

SDRtrunk-parity subsystems. Outbound call streaming (Broadcastify Calls / RdioScanner / OpenMHz / Icecast), wideband baseband recording + offline replay, the GPS / location and affiliation subsystems, the decoded-message log, and per-talkgroup stream / record / mute / icon policy all ship and are covered by the test suite. Analog FM trunking (Motorola Type II, EDACS, LTR, MPT 1327) decodes voice through the composer’s FM chain.

Remaining gaps

Additional SDR hardware

RTL-SDR, HackRF (One / Jawbreaker / Rad1o / Pro), Airspy R2 / Mini, and Airspy HF+ (Discovery / Dual Port / legacy) are all supported by pure-Go drivers with mock-transport unit tests. HackRF (Pro board-ID detection, fpga_dc_block, dc_avoid, rf_amp) and Airspy (macOS async bulk-IN rework, native-rate behaviour) have been exercised and fixed against attached hardware in the field, and all three backends have hardware-gated harnesses (internal/sdr/{airspy,airspyhf,hackrf}/*_real_test.go, gated on GOPHERTRUNK_{AIRSPY,AIRSPYHF,HACKRF}_REAL; make test-*-real). Remaining: Airspy HF+ has had no attached-hardware exercise — its harness has never been run against a real unit. SDRPlay / USRP / BladeRF have no local zero-CGO driver (their vendor libraries are C), but are reachable over SoapyRemote — USRP X310 and B210 rigs are field-tested that way.

Digital-voice composer chains

FM (incl. analog trunking), DMR, P25 Phase 1 / 2, TETRA TMO + DMO (clean-room ACELP), NXDN, dPMR, and D-STAR decode to audio. TETRA voice is verified bit-exact against the ETSI EN 300 395-2 reference codec (via the env-gated harness in internal/voice/acelp/etsi_reference_test.go — the ETSI vectors are copyrighted and not committed). NXDN, dPMR, and D-STAR are wired end-to-end through the composer but not yet verified on air: each chain’s AMBE interleave table is a documented placeholder awaiting a real voice capture (internal/radio/{nxdn,dpmr,dstar}/voice_ambe.go — dPMR anchors on the FS1/FS2 voice syncs and renders through the AMBE+2 3600x2450 “ambe2-dmr” decoder; D-STAR anchors its 96-bit DV cadence on the Slow Data sync and renders through the base AMBE 3600x2400 “ambe2” decoder). TETRA DMO (direct mode, protocol: tetra-dmo) records audio through the same ACELP vocoder but is experimental: call source / destination identity is not decoded (recordings file under group 0) and the chain still awaits its on-air A/B (issue #1003). YSF voice and EDACS ProVoice are still bypassed — their calls are followed and logged but not yet turned into PCM.

Per-protocol on-air FEC inner layers

Every protocol’s ControlChannel.Process adapter ships a working IQ → CC chain. The spec-correct chain is on by default for every protocol; operators with pre-stripped capture files opt out per-system. See opt-in-features.md for the full table.

The inner FEC layers still pending real-air validation:

  • NXDN per-protocol interleaver + puncture. ViterbiSpec mode runs the full §4.5.1.1 chain; ViterbiOn is the simpler bare-bones path the older MMDVMHost / DSDcc fixtures use. Both are wired through the connector. Calibration against captured MMDVMHost transmissions is the step that lands next. The 4-FSK slicer’s peak-deviation reference surfaces as a per-system nxdn_deviation_hz knob (default 1800 Hz per the Common Air Interface). The skip-gated real-air harness at cmd/gophertrunk/integration_cc_nxdn_realair_test.go runs acceptance criteria automatically once a contributor drops a .cfile + .metadata.json pair into samples/nxdn/.
  • TETRA on-air recovery margins — now characterised (TMO). A real marginal-signal control-channel capture is committed at internal/scanner/ccdecoder/testdata/tetra_cc_sync_loss_2s_144k.cs16 and pinned in CI by internal/scanner/ccdecoder/pipelines_tetra_equalizer_test.go: the marginal regime (~10 dB in-channel SNR) decodes ~12% of its BSCH without equalization and ~100% with the blind SnapshotCMA equalizer, and soft-decision TCH/S decoding lifts CRC-valid voice yield ~1.9× across the reporter’s captures. What remains TETRA-side is the DMO on-air A/B (issue #1003) and a validating .metadata.json sidecar for samples/tetra/ so the skip-gated TMO real-air harness runs.
  • P25 Phase 2 traffic-channel MAC descramble (issues #915, #773, #451). The superframe now locks on real air under any dibit rotation (the differential-H-DQPSK residual-carrier ambiguity fixed in #943), so the ISCH classifies the MAC sub-frames correctly and mac_pdus > 0. What still does not work is recovering a valid MAC PDU from the payload: mac_rs_valid stays 0, so the clear-MAC source RID (#915) and talker alias (#773) never surface. Replaying the reporter’s Victorian MMR (WACN 0xBEE00) Phase 2 voice .cfile offline ruled out — against real bytes — every tractable hypothesis: the payload’s trellis path-metric is ~42/73 (random) undescrambled while the un-scrambled sync + ISCH decode clean, so the payload is transformed; but descrambling it (both after the trellis at per-slot offsets {index·360, index·144, 0} and before the trellis at offsets {0, superframe index·360+64}) across all 16.7 M (WACN, SysID, NAC) identity seeds never validates the outer RS and only nudges the best trellis metric 41.8 → 33.8 (still garbage). A seed-independent structure test (XOR two same-slot payloads, which must share the sequence under any per-superframe-restarting PN44 model — the sequence then cancels and the linear trellis code makes the XOR a clean codeword) stays at ~41 error, so the transform is not a fixed positional PN44 XOR keyed by the network identity. What remains is a continuous / counter-seeded scrambler or a MAC FEC chain (trellis params / interleaver / RS-before-conv order) different from what GT models — pinning it needs the TIA-102.BBAC §7.x scramble/FEC definition or an SDRtrunk / OP25 P25P2 reference to cross-check. The site’s real System ID + NAC would remove the seed unknown and separate a seed-formula bug from a scramble-structure bug. The mac_rs_valid census counter (#934) is the before/after metric once a candidate mapping is tried, and the RS-valid gate keeps the mis-decoded bytes from injecting a bogus source RID in the meantime.

  • DMR 2-slot interleaved voice — now the Tier II conventional & Tier III default (issue #644). A DMR carrier is 2-slot TDMA, so a real outbound stream interleaves both timeslots’ bursts. The single-slot voice.NewDecoder splices the two slots together into garbled, encrypted-sounding audio — the #644 report. voice.NewInterleavedDecoder decodes the carrier correctly: it locks each slot’s burst A on its own voice sync and gathers that slot’s B–F by the same-slot stride, emitting one superframe per slot tagged by VoiceSuperframe.Phase. It now auto-detects the on-air same-slot cadence per call — 264 dibits (no inter-burst CACH) vs 288 (a 12-dibit CACH precedes each burst on live BS-sourced outbound air) — by slicing bursts B–E at each candidate and locking onto the one that reassembles a CRC-valid embedded Link Control (a wrong cadence cannot). The decoder also surfaces that LC’s talkgroup + source on VoiceSuperframe.LC, so a phase binds to a concrete talkgroup — the absolute TS1/TS2 label the identical BS-sourced burst-A sync cannot give. The chain is unit-tested against synthetic interleaved + embedded-LC vectors at both cadences (TestInterleavedDecoderAutoDetects{CACH,NoCACH}Cadence). A cadence chosen only by FEC quality (no LC yet) is held provisionally: a later CRC-valid LC — or a clear FEC winner at a different stride — overrides a wrong early guess and re-locks the correct cadence, so one bad guess no longer garbles the rest of the call (TestInterleavedDecoderLCOverridesWrongProvisionalCadence). The interleaved + LC path is the default for DMR Tier II conventional and Tier III (#644, extended to Tier II after a field report of garbled “DJ-scratch” audio on a dmr-tier2 site — the same single-slot/2-slot mismatch): the daemon tags those systems’ voice grants so the composer runs NewInterleavedDecoder and routes each call to its timeslot by the embedded LC’s talkgroup (a slotRouter). DMR Tier I is direct-mode simplex (genuinely single-slot) and stays on NewDecoder. dmr_interleaved_voice is a tri-state override (unset = protocol default; true/false to force). One piece still wants a real IQ capture to cross-check: the exact ETSI embedded-signalling de-interleave order, the EMB QR(16,7) FEC (read systematically for now), and the 5-bit CRC polynomial — currently internally consistent (encode↔decode round-trip) but not yet validated against captured traffic. The skip-gated harness internal/voice/composer/dmr_2slot_realair_test.go (run with -tags integration and GOPHERTRUNK_DMR_2SLOT_CFILE) is where a contributor drops a real capture to confirm those remaining constants. Because that embedded LC is capture-pending, the slotRouter no longer hard-drops a call when the LC never decodes: a matching LC still binds (and corrects) the slot, but after a short grace window with no LC it falls back to the active slot’s phase so audio still records instead of producing empty files (#644). The decode-quality log reports lc_superframes and notes once when a call records via the phase fallback, so a capture that exercises the fallback is easy to spot.

  • AMBE+2 synthesis parity with IMBE (#644 follow-up). Once the timeslot fix made DMR speech intelligible it sounded metallic (“tin can”) — the buzz from fully phase-coherent voiced synthesis. The AMBE+2 decoder (DMR plus P25 Phase 2 / NXDN / dPMR) now runs the same three post-synthesis stages the IMBE decoder already had: §6.3 voiced-phase regeneration (mbe.SynthVoicedDispersed, the de-buzz, scaled by the unvoiced-harmonic fraction), a DC-removal high-pass (mbe.DCBlock) ahead of the AGC, and error-rate adaptive smoothing (mbe.Smoother). The DMR voice chain forwards the per-frame Golay corrected-bit count (errAwareRawSink → voice.ErrorAware.SetFrameErrors) to drive the smoother. Clean fully-voiced frames are bit-identical to before; only the metallic timbre changes.

Digital-voice level calibration

Pure-Go IMBE / AMBE+2 emit real audio end-to-end. The comparison harness at internal/voice/calibrate/ (CLI: cmd/voice-calibrate) is ready, and the AMBE+2 capture fixture (internal/voice/ambe2/testdata/dmr-voice.raw) is committed. Still missing are the DSD-FME / OP25 reference WAVs — internal/voice/imbe/testdata/p25-p1-voice{.raw,-dsdfme.wav} and internal/voice/ambe2/testdata/dmr-voice-dsdfme.wav — which also gate the final quality sign-off of the default-on spec-faithful §6.2 spectral-amplitude enhancement (recordings.spec_amplitude_enhance). Knox / call-alert AMBE+2 tones (b₁ ∈ [144, 163]) are vendor-specific and stay silent until per-vendor frequency tables land; operators with a curated table register it via ambe2.RegisterPreset. See vocoders.md for the licensing posture and sourcing checklist.

Recently-shipped items live in CHANGELOG.md; near-term plans live in Roadmap.