P25 End to End, Part 12: The Weak-Signal Gap — P1 Voice's Missing Levers

By Matt Cheramie · September 14, 2026

Part 12 of P25 End to End, a 14-part deep dive that follows North America’s dominant trunking protocol through GopherTrunk — from a raw C4FM carrier to recorded, named, multi-site voice. Part 11 watched a whole system through one wide capture and met the twin pair that let a fix land on one side and miss the other. This part is the one the opener promised: the honest inventory. On a marginal Phase 1 voice call, a hardware radio decodes clean audio while GopherTrunk recovers a handful of frames — and we know exactly which two stages are missing, why they haven’t been built, and what single recording unblocks them.

TL;DR: The default P25 Phase 1 C4FM voice receiver (internal/radio/p25/phase1/receiver/receiver.go) is FM discriminator → spec matched filter → CoarseAFC → Mueller-Müller timing → symbol AGC → 4-level slicer → hard Golay/Hamming IMBE FEC — no channel equalizer, no soft-decision voice FEC. The synthetic sweep (TestSweepImplementationLossBudget) puts a number on the fragility: at 1% symbol-error rate the C4FM path needs ~23.8 dB more Es/N0 than the coherent 4-PAM reference, while the CQPSK twin — which carries a T/2 blind equalizer (the fse field in cqpsk.go) — sits ~3.85 dB off coherent QPSK. The missing levers are the pair that roughly doubled TETRA’s marginal yield; the port is gated on a real weak C4FM voice capture, and samples/p25/README.md + TestReplayP25RealCaptureMetrics are the baseline instrument waiting for it.

Key takeaways

• The gap is structural, and it is measured. Every stage of the C4FM voice chain works as designed; it simply lacks the equalization and soft-FEC stages the marginal regime demands, and the BER sweep quantifies the headroom against a reference the CQPSK path tracks within 4 dB.
• This is the twin-path thread inverted. Usually a fix lands on one twin and misses the other; here a capability — the blind equalizer — was only ever built on the CQPSK/LSM twin and on Phase 2.
• The soft machinery is half-installed. The C4FM receiver already emits two log-likelihood ratios per dibit (BitLLRSink) and the control channel spends them on soft TSBK decode — but the IMBE Golay/Hamming decoders never see a soft bit.
• No fix without a capture. A green synthetic is not proof of an on-air fix — the #764/#771 lesson. The highest-leverage contribution to GopherTrunk’s P25 support right now is a recording, not a patch.

Cheat sheet

Concern	What it does	Where it lives
C4FM voice chain (all hard)	discriminator → MF → AFC → M&M → AGC → slicer	internal/radio/p25/phase1/receiver/receiver.go (DemodC4FM branch)
The equalizer next door	T/2 fractionally-spaced blind CMA on the linear path	internal/radio/p25/phase1/receiver/cqpsk.go (fse *equalizer.FSE)
Soft plumbing, half-installed	per-dibit LLRs feed soft TSBKs, not IMBE	receiver.go (BitLLRSink) → control.go (StashSoft)
Measured fragility	C4FM ~23.8 dB off coherent 4-PAM at 1% SER	receiver/sweep_test.go (TestSweepImplementationLossBudget)
Baseline instrument	EVM / SNR / FSW-margin / NID / TSBK from a real capture	cmd/gophertrunk/p25_realcapture_metrics_test.go
The ask	a weak C4FM voice .cfile + metadata sidecar	samples/p25/README.md (weak-signal voice section)

In this post

• The report — a hardware Astro Spectra sets the bound.
• The chain, audited — what the C4FM voice path runs and discards.
• The twin that got the equalizer — fse, Phase 2, the inverted thread.
• The gap, measured — what the sweep’s loss budgets actually say.
• Why the port is parked — proven levers, missing evidence.
• The instrument is waiting — what to record, and what happens next.

The report: a hardware radio sets the bound

The report that defines this part is a controlled experiment an operator ran without meaning to. A marginal P25 Phase 1 voice call decodes cleanly on a hardware Astro Spectra; the same call, on the same antenna, through GopherTrunk yields about 4–5 IMBE frames from a whole transmission. Not silence, and not a failure to lock — a lock that extracts almost nothing, the signature of the marginal regime in its purest form.

The hardware radio matters because it bounds the problem: the information survived the antenna, so whatever is lost is lost after the ADC. There is a telling brush in the numbers, too — an LDU carries 9 IMBE frames in 180 ms, so 4–5 frames is under one LDU, well short of the minAutotuneLDUs = 5 threshold in internal/voice/composer/p25p1_voice.go that skips the autotune measurement for calls too short or noisy to trust. The guard is correct; that real calls keep tripping it is the indictment.

Commercial subscriber radios earn their weak-signal margin with two tools: adaptive equalization and soft-decision error correction. Which does the default C4FM voice path run?

The chain, audited

The receiver’s own doc comment is the inventory:

// internal/radio/p25/phase1/receiver/receiver.go (shape) — the C4FM branch
//	DemodC4FM (default):
//	  IQ
//	    → FM discriminator (internal/dsp/demod.FM)
//	    → spec P25 C4FM matched filter (demod.P25C4FMRxTaps — not RRC)
//	    → coarse AFC: residual carrier-offset removal (demod.CoarseAFC)
//	    → Mueller-Müller symbol clock recovery (sync.MuellerMuller)
//	    → 4-level slicer → C4FM symbol → 0..3 dibit

Downstream, the LDU voice path (Part 8) runs ExtractVoiceFrames: hard Golay and Hamming decoding per IMBE subframe. Audited against the marginal regime:

• No equalizer, anywhere on the C4FM path. No stage between timing recovery and the slicer inverts inter-symbol interference; a multipath-smeared 4-level eye goes into symbol decisions smeared.
• A hard slicer feeding hard FEC. The slicer collapses each symbol to a dibit and throws away its distance from the thresholds — the confidence a soft FEC would spend. Golay/Hamming then defend a coin-flip bit as firmly as a certain one.
• The soft machinery exists — for the control channel. The C4FM receiver already derives two log-likelihood ratios per dibit (Options.BitLLRSink — sign-axis distance for the high bit, inner/outer-threshold distance for the low), and the control channel’s StashSoft path spends them on a per-bit soft TSBK Viterbi behind the p25_phase1_soft_decision config key. The plumbing runs right past the voice path: IMBE’s Golay and Hamming decoders take hard bits, full stop.

Nothing in that list is a bug. It is a fair-weather design meeting a use case that is not fair weather.

The twin that got the equalizer

This series’ running thread — P25 as a family of twins, each pair a place fixes drift apart — arrives here inverted: not a fix that missed one twin, but a capability that only ever landed on one.

Part 6 covered the linear path in full: the opt-in DemodCQPSK receiver carries a T/2 fractionally-spaced blind equalizer, and its configuration is worth re-reading with this part’s eyes:

// internal/radio/p25/phase1/receiver/cqpsk.go (shape)
const (
    cqpskFSESymbolSpan = 6     // 12 T/2 taps, centre tap on-time
    cqpskFSEStep       = 0.025 // brisk CMA step — short captures must converge
    cqpskFSELeak       = 5e-4  // pulls taps toward identity on clean input
)
/* … */
fse: equalizer.NewFSE(cqpskFSESymbolSpan, cqpskFSEStep, 1.0, cqpskFSELeak),

The comments above those constants record why fractional spacing matters: a T/2 equalizer synthesizes the receive matched filter implicitly, opening both simulcast multipath ISI and pulse-shape mismatch — things a symbol-spaced equalizer cannot (issue #492, told in four acts). P25 Phase 2 wires an equalizer as well (Part 7). Default C4FM Phase 1 voice is, literally, the odd path out.

Side by side: the C4FM voice chain versus the TETRA voice chain. Every solid block on the P25 row works; the two dashed blocks are the diagnosis — and both already exist elsewhere in the tree.

The gap, measured

“Noise-fragile” is not a vibe here; it is a committed CI number. The synthetic BER-vs-SNR sweep in receiver/sweep_test.go (TestSweepImplementationLossBudget) drives the real receiver on both demod paths across an injected-SNR ladder and gates the result against closed-form references:

Path	Reference curve	Measured loss @ 1% SER	Committed budget
CQPSK (DemodCQPSK)	coherent QPSK (metrics.SERCoherentQPSK)	~3.85 dB	6.0 dB
C4FM (DemodC4FM)	coherent 4-PAM (metrics.SER4PAM)	~23.8 dB	27.0 dB

Read the two numbers together. The linear path — coherent demod, equalizer, Gardner timing — reaches 1% symbol errors within ~4 dB of the coherent QPSK bound. The C4FM discriminator path needs ~24 dB more Es/N0 than coherent 4-PAM. The honest caveat is in the test’s own comment: coherent 4-PAM is a bound no FM discriminator reaches, and the budget “is a regression ceiling … NOT a claim that the gap is small” — part physics of non-coherent detection, part timing-slip behaviour the measurement folds in. But the ceiling stops the path silently getting worse, and it quantifies the headroom that justifies demod work. Twenty-plus dB of headroom is why a subscriber radio and GopherTrunk disagree so completely about the same marginal call.

Why the port is parked

Both missing stages are proven technology in this codebase. The blind snapshot equalizer and soft-decision decoding (Weak-Signal Engineering Parts 4 and 8) roughly doubled TETRA’s CRC-valid voice yield across six operator captures (410 → 778 soft-decision TCH/S, one call going 4 → 207), and the soft path recovered the ~70% of a marginal call’s bursts the hard gate dropped. The physics doesn’t care which trunking protocol rides on top.

So why not port it this afternoon? Because the project has been burned, twice, by exactly that move: issue #764 was closed twice on fixes that were green against synthetic fixtures while the live symptom persisted (issue #771 is the receipt). A synthetic channel exercises the impairments you imagined; the operator’s terrain exercises the ones you didn’t. The standing gate that came out of it: a weak-signal fix lands only with a baseline on a real capture and an A/B where decode yield is the verdict — never EVM, which has lied to this project before. The sibling series covers this path from the method’s side (The Odd Path Out); this series adds the protocol context: it is the last unequipped twin.

The instrument is waiting

Everything buildable without the capture is built. TestReplayP25RealCaptureMetrics (cmd/gophertrunk/p25_realcapture_metrics_test.go, tag integration) replays any .cfile dropped into samples/p25/ through the production receiver and reports pre-FEC EVM, estimated SNR, the FSW sync-margin distribution, and NID/TSBK yields — every metadata bound optional, so a capture can be measured before anyone commits to a pass/fail threshold. The pipeline is validated end to end: a live UHF C4FM control-channel capture (449.875 MHz, NAC 0x2C1) grades at EVM ≈ 12.7%, SNR ≈ 14.5 dB, NID trusted 31 / failed 0, TSBK decoded 36 / CRC-failed 0.

What’s missing is the voice capture. samples/p25/README.md spells out the ask: tune the granted voice frequency during a call that is genuinely weak or multipath-degraded — a strong, clean call cannot exercise a missing equalizer — note the hardware radio’s contemporaneous decode in tool_cross_check, and drop the .cfile + .metadata.json pair in (the binary stays git-ignored; the sidecar is committed). From there the sequence is mechanical: baseline, port one lever — the capture’s own numbers decide whether ISI says equalizer-first or a clean eye with FEC failures says soft-first — and A/B the LDU/IMBE yield.

Where this goes next

A series that leans this hard on “the test that passes because both sides share the same bug” owes you the full testing story. Part 13 climbs the whole pyramid — literal byte vectors, impairment-swept synthetic streams, committed captures, on-air verification — using P25’s own bugs as the case studies.

FAQ

Why can a hardware scanner decode a P25 call GopherTrunk can’t? Because commercial subscriber hardware ships the two stages this path lacks: adaptive equalization ahead of the symbol decisions and soft-decision error correction behind them. On clean signals the difference is invisible; in the marginal regime those stages are worth many dB of effective sensitivity.

Is the P25 control channel affected the same way? Less so. The control channel can opt into per-bit soft TSBK decode (p25_phase1_soft_decision), fed by the same BitLLRSink LLRs the voice path ignores — and TSBKs repeat, so a missed block usually costs nothing. IMBE frames don’t repeat: every hard-FEC failure is 20 ms of audio gone, which is why the operator-visible gap is voice.

Couldn’t I just run the CQPSK receiver on the weak C4FM site? No. The equalizer lives on a linear demod chain that expects an LSM-style phase trajectory, and the two paths are not interchangeable (the LSM myth postmortem covers why; issue #935 covers when each applies). The equalizer has to come to the C4FM chain, not the signal to the equalizer.

What exactly should I record to help? A P25 Phase 1 C4FM voice-channel IQ recording of a marginal call — the granted frequency, not the control channel — as a .cfile at ≥ 48 kHz with a metadata sidecar per samples/p25/README.md, ideally noting what a hardware radio made of the same call. That one file converts a parked fix into a measurable A/B.

Does the 23.8 dB figure mean the C4FM demod is broken? No — the reference is coherent and the demod is not, plus real implementation loss on top. The number’s job is to be a ceiling (the gate fails past 27 dB) and a motivator: the measured size of the pool the equalizer and soft FEC would fish in.

Series navigation

Part 12 of 14 · ← Part 11: Wideband P25 — Watching the Whole System at Once · Next → Part 13: Testing P25 Without a Tower