The Operator's Cookbook, Part 12: Two Antennas, One Signal — A Diversity Build

By Matt Cheramie · September 14, 2026

Part 12 of The Operator’s Cookbook, a 14-part series of complete, copy-paste GopherTrunk builds — one working rig per part, antenna to browser. Part 11 put the rig in a closet and taught it to survive unattended. This part is the advanced RF build: two antennas feeding one decoder, phase-coherently combined so a fade on one branch is covered by the other. Diversity is the one upgrade that can silently do nothing — or, before the aligner, actively hurt — so the log lines that prove it’s working matter more here than anywhere else in the series.

TL;DR: Diversity needs one radio with two coherent RX channels — a USRP B210 (shared LO) or an X310 with TwinRX daughterboards (independent PLLs) — mounted over sdr.soapy_remote[] with diversity: "mrc" and antenna: [RX1, RX2]. Health is one log line every 30 s: soapyremote: MRC diversity branches, and branch_phase_deg is the instrument: constant ⇒ shared-LO hardware, mrc-static is fine; walking ⇒ independent PLLs, you want tracking mrc. The honest ceiling: MRC is guaranteed ≥ the best branch alone, but a real gain over it only shows on weak signals — prove yours with a diversity_capture pre-combine dump and TestDiversityCombinerReplay, scored by decode yield, never by dBFS.

Key takeaways

• Diversity is the last lever, not the first. A better antenna, feedline and gain staging (Parts 7–9 of The Analog Edge) buy more margin per dollar. Build this when one antenna genuinely fades.
• Hardware class decides the mode. A shared-LO front end has one fixed inter-branch phase; independent PLLs lock at a random phase that then walks. branch_phase_deg tells you which class you own, before any capture.
• Coherence, not dBFS, is the verdict on the combine. The coherence figure is independent of RF gain — raising gain to make diversity “engage” is never the answer, and the config file says so in as many words.
• MRC ≥ best branch is a floor, not a promise of gain. On a strong signal every combined arm just matches the better antenna. The gain case lives on weak signals, and the pre-combine capture A/B is how you check yours.

Cheat sheet

Concern	What it does	Where it lives
Enable combining	open RX0+RX1, combine into one stream	sdr.soapy_remote[].diversity: "mrc"
Freeze the estimate	one-shot calibration, for A/B or shared-LO rigs	diversity: "mrc-static"
Antenna ports	per-channel port assignment, validated + read back	antenna: [RX1, RX2]
Hardware-class instrument	constant vs walking inter-branch phase	branch_phase_deg in the MRC health line
Pre-combine evidence	per-branch IQ dump only the driver can make	diversity_capture, diversity_capture_seconds
Offline A/B	replay both branches through every combiner arm	TestDiversityCombinerReplay (GT_DIVERSITY_CAPTURE)
The concepts	why MRC works and where it can’t	antenna diversity, MRC gotchas

In this post

• What you’re building — two antennas, one coherent front end, one combined stream.
• The shopping list — the two hardware classes, honestly priced.
• The config — diversity: mrc on a SoapyRemote device, every key verified.
• First run — what healthy looks like — the MRC health line, field by field.
• The honest ceiling & the pre-combine A/B — what MRC can buy, and how to prove it.
• When it doesn’t work — symptom → cause → fix, then variations.

What you’re building

The Part 8 remote-radio build with a second antenna: a two-channel SDR runs SoapySDRServer near the antennas, GopherTrunk opens both RX channels, and the driver phase-aligns and sums them into one maximised-SNR stream before the decoder ever sees IQ. Everything downstream is unchanged; diversity is invisible except in the health log and, when the channel fades, in the calls that keep decoding.

Three driver stages do the work, in order: a pre-combine capture tap (raw, so evidence stays untouched), an inter-branch aligner that removes the fractional-sample start skew between the streams (on one field rig, 2.6 samples of skew made the combine decode 22% worse than the best branch alone — the aligner is why that can’t happen anymore), and the MRC combiner, which estimates one complex gain per calibration window and either tracks (mrc) or freezes it (mrc-static). How each stage earned its place is Weak-Signal Engineering Parts 10–11; this recipe is the operator’s side.

The decoder never knows there were two antennas — the capture tap, aligner and combiner live inside the driver, and one log line reports whether they're earning their keep.

The shopping list

The honest first line: most rigs don’t need this. You need one radio with two RX channels sharing a frequency reference — two separate dongles will not do.

Item	Class	Notes
USRP B210 (or similar 2-ch shared-LO SDR)	shared LO	one chip, both channels — inter-branch phase is a hardware constant; mrc-static territory
USRP X310 + 2× TwinRX	independent PLLs	separate daughterboards (rx_subdev_spec=B:0 A:0): frequency-locked, but the relative phase is random per lock and walks after it — tracking mrc territory (USRP notes)
Two antennas, same band + polarisation	—	co-located — a wavelength or two apart, not metres — with decent feedline
A host for SoapySDRServer	—	the Part 8 remote pattern; wired network, budget for two channels of CS16

No external reference oscillator is required for the X310/TwinRX class: a field capture proved the dual-daughterboard stream coherent as delivered (wideband coherence 0.95+ on a healthy run) — handling the phase is the calibrator’s job.

The config

The Part 8 SoapyRemote block plus three diversity keys, all verified against config.example.yaml:

sdr:
  sample_rate: 250_000        # soapy_remote isn't bound to RTL rate limits
  soapy_remote:
    - addr: "192.168.1.60:55132"
      driver: "uhd"
      args: "rx_subdev_spec=B:0 A:0"   # X310: one channel per TwinRX
      serial: "usrp-attic"
      role: control
      format: "CS16"
      gain: "auto"
      diversity: "mrc"                 # "" | "mrc" | "mrc-static"
      antenna: [RX1, RX2]              # RX1→channel 0, RX2→channel 1
      # diversity_capture: "../iq/mrc/precombine"
      # diversity_capture_seconds: 30

trunking:
  systems:
    - name: "Metro-TETRA"
      protocol: tetra
      control_channels:
        - 467_912_500

Three notes. antenna: is the only correct way to pick ports — names are device-specific (a B210 has TX/RX and RX2, a TwinRX has RX1 and RX2), and GopherTrunk validates the list and reads it back, so a config moved between rigs fails loudly. diversity: "mrc" re-estimates the branch gain continuously; "mrc-static" freezes it after one estimate — the first run tells you which to pick. And the combine happens below the protocol layer, so a Part 4 TETRA rig or a Part 1 P25 rig rides it identically.

First run — what healthy looks like

Start the daemon and let it lock as usual (tetra cc locked freq=467912500 mcc=… mnc=… here). The diversity heartbeat is one INFO line every 30 s:

INF soapyremote: MRC diversity branches addr=192.168.1.60:55132 branch_dbfs="ch0=-51.2 ch1=-50.6" reference_branch=1 calibrated=true coherence=0.95 branch_gain_db=-1.2 branch_phase_deg=145.3 mode=mrc updates=6682 holds=0

Read it field by field — this line is the whole cockpit:

• branch_dbfs — both branches alive and within a few dB. A branch ~10 dB down is a weak antenna or feedline, and MRC on a floor-limited branch is roughly no-gain: fix that branch’s RF first.
• coherence — the normalised cross-correlation between branches, gain-independent by construction. A healthy co-located pair reads 0.9+ at moderate bandwidth; on wide captures lower numbers are normal and the lock gates scale with the window — coherence-not-dBFS is the full story.
• branch_phase_deg — the no-capture hardware-class instrument. Constant across lines means shared-LO and mrc-static would serve; walking (a TwinRX rig measured about −0.2°/s — a frozen constant decays over minutes) means independent PLLs and tracking mrc.
• updates / holds — accepted vs rejected calibration windows. Healthy is updates climbing with holds near zero; updates frozen while holds climbs means coasting on a stale gain, and it WARNs after 90 s.

One more line to know on sight, once per stream or retune — the aligner latching the start skew (per-stream, which is why it’s re-measured every time rather than calibrated once):

INF soapyremote: MRC inter-branch delay measured — delaying the early branch to align the combine delay_samples=0.41 peak_rho=0.94

The honest ceiling — and the A/B that tests it

Now the part a recipe owes you before you spend this money. Post-aligner, MRC is a no-harm combine: every capture A/B run to date scores the combined arms within one decoded frame of the best branch alone. But those captures decode at their ~100% yield ceiling — matching the best branch is the ceiling there — so a real gain over the best branch is still undemonstrated on air. Theory says it shows on weak, fading signals; a weak-signal pre-combine capture where the combined arm beats both branches is genuinely wanted upstream.

Second limit: the combine applies one complex gain to the whole wideband stream. Antennas metres apart give every carrier its own phase difference, and a single scalar can only align one — the signature is coherence stuck around 0.3–0.5 that no tracking improves. Co-locate the antennas; per-channel combining after the DDC is future work, not a config option (MRC gotchas).

The pre-combine A/B

Every other IQ tap — baseband.auto_record, the scope feeds — sits downstream of the combiner, so a capture from any of them has one combiner baked in. diversity_capture is the one tap that records the branches raw, straight after de-interleave:

INF soapyremote: diversity capture armed — dumping pre-combine per-branch IQ prefix=../iq/mrc/precombine seconds=30 branches=2
INF soapyremote: diversity capture complete — replay it with GT_DIVERSITY_CAPTURE to A/B combiners offline

You get <prefix>.br0.cs16, <prefix>.br1.cs16 and a .diversity.json sidecar; a datagram that didn’t carry every branch is dropped from both files, never written short — one short write silently desynchronises the pair. Then grade every combiner against your own air:

GT_DIVERSITY_CAPTURE=../iq/mrc/precombine.diversity.json \
  go test ./cmd/gophertrunk -run TestDiversityCombinerReplay -v

The harness prints a windowed coherence/gain/phase trace with plain-text verdicts (“branch phase is essentially CONSTANT” vs “branch phase WALKS”), measures the inter-branch delay, and decodes the capture through each branch alone plus static, tracking, aligned and narrowband combine arms — scored by CRC-clean decode yield, the one metric that can’t flatter a combiner. That table answers the mode question with evidence instead of folklore.

When it doesn’t work

Symptom	Likely cause	Fix
WARN MRC diversity got 1 of 2 channels in the stream	the server never delivered the second RX channel	check args (rx_subdev_spec) and that the device really has two RX channels
WARN MRC diversity branch is dead	that antenna’s connection, feedline or per-branch gain	reseat/re-test that branch’s RF path; port assignment is antenna:
WARN MRC diversity branches are not coherent	different bands/polarisation, widely separated antennas, or no frequency lock	co-locate, match polarisation, check clock_source/time_source; a far-down branch_dbfs wants its gain staging fixed
Calibrated once, then WARN has not accepted a calibration window	the wideband-scalar limitation — one gain can’t align every carrier from separated antennas	move the antennas together, or freeze deliberately with mrc-static
Combine decodes worse than one antenna alone	not skew (the aligner handles it) — a branch is dragging the estimate	run the pre-combine A/B; the arm table names the culprit
Tempted to raise gain until diversity “engages”	the old absolute-power trap	don’t — coherence is gain-independent (gain staging)

That last row has history: an earlier build gated calibration on a fixed coherence constant, and its WARN told a bandwidth-diluted operator that raising RF gain would not help — exactly wrong in their regime (their weak branch needed +5 dB of per-branch gain). The gates now bound the estimate’s phase error instead. The series-wide lesson: never trust an absolute-dBFS rule; trust coherence and decode yield.

Variations

• diversity: "mrc-static" — one-shot calibration, frozen. Correct for shared-LO hardware, and the standard A/B reference arm on any rig.
• Longer evidence — diversity_capture_seconds accepts up to 1200 s (a 1 GiB/branch cap always applies); at 200–250 kS/s two CS16 branches are only ~1.6 MB/s total, so take the long capture.
• Single-channel, port-pinned — diversity: "" with antenna: [RX2]: a plain remote SDR with the port explicit and validated — the “each branch alone” baseline before you commit.
• Voice under diversity — the combined stream is an ordinary tuner to the daemon; role: wideband with voice_taps on top works exactly as in Part 1.

Where this goes next

The rig now has every piece of hardware this series will ask you to buy — but every talkgroup and radio is still a number. Part 13 is the zero-hardware recipe: alias CSVs and their real columns, naming things live from the browser, and exporting the result back to files you own.

FAQ

Do I need a diversity setup for a scanner rig? Almost certainly not first. It’s the only lever that costs serious hardware, and it addresses one regime: a signal that fades at your location. Antenna, feedline and gain staging buy more margin per dollar; build this when those are done and a marginal system still swings.

Should I use mrc or mrc-static? Let branch_phase_deg answer: run mrc and watch the health line for a few minutes. Constant phase means shared-LO hardware — mrc-static is equivalent and simpler. Walking phase — TwinRX-style independent PLLs — means stay on mrc, because a frozen constant decays over minutes.

Can I do diversity with two RTL-SDR dongles? No. The combine needs two RX channels behind one clock in one sample stream — diversity: mrc opens RX0+RX1 on a single SoapyRemote device. Separate dongles have independent oscillators and USB timelines; there is no constant phase for the calibrator to find.

Why is my coherence only 0.3 when everything decodes fine? Wideband coherence is diluted by every hertz of noise-only bandwidth around your carrier — a clean channel occupying a small fraction of a wide capture reads low even when the branches agree perfectly in-channel, which is why the lock gates scale with the estimation window. Judge the combine by decode yield and the WARN lines.

Series navigation

Part 12 of 14 · ← Part 11: The Closet Appliance — Pi, systemd & Docker · Next → Part 13: Naming Everything — Aliases, Labels & Exports