Working SDR setup and decoded results for Baton Rouge, LA 70809. Date: 2026-08-30.
- LWIN P25 trunking — fully decoded. Louisiana's statewide P25 system, Baton Rouge Simulcast site. System identity, site topology, and live talkgroup traffic recovered off the air and cross-verified against RadioReference.
- FM RDS — decoded, including live now-playing song metadata.
- Complete LWIN reference database pulled locally (
reference/).
NAC 0x1BD
WACN 0xBEE00 <- Louisiana statewide
SYSID 0x1BD
RFSS 1 Site 13 -> "Baton Rouge Simulcast" (East Baton Rouge Parish)
Control channel: 773.05625 MHz (op25 self-reported ch1: 16e8 -> 773.056250)
From a 40 s capture: 1599 TSBK messages, 0 discarded frames.
| Site | Hex | NAC | Name | Parish |
|---|---|---|---|---|
| 10 | 0x0A | — | Geismar | Ascension |
| 13 | 0x0D | 1BD | Baton Rouge Simulcast | East Baton Rouge (receiving) |
| 17 | 0x11 | 1B0 | Denham Springs | Livingston |
| 21 | 0x15 | 1B8 | Saint George | East Baton Rouge |
| 28 | 0x1C | — | Baywood | East Baton Rouge |
| 30 | 0x1E | 1B4 | Ramah | Iberville |
| 31 | 0x1F | 1BF | Zachary | East Baton Rouge |
| 32 | 0x20 | — | Sage Hill | West Feliciana |
| 37 | 0x25 | 1B1 | Livingston | Livingston |
| 44 | 0x2C | — | Clinton | East Feliciana |
| 50 | 0x32 | 1BD | South Baton Rouge | East Baton Rouge |
| 72 | 0x48 | — | Livonia | Pointe Coupee |
24 FREQS · 4 CONTROL — control: 773.05625, 774.54375, 851.0375, 851.4875
Notable — site 13 is split across two bands. The control channel is at 700 MHz
(773.05625, +28 dB here) but voice traffic is granted onto 800 MHz (851-853 MHz):
op25 logged new freq=852.150000, 851.837500, 852.987500, 852.750000, 852.562500.
An early 850-870 MHz sweep with the old antenna found nothing, which was misleading on two counts: the antenna has since been swapped, and voice channels are bursty — a time-averaged sweep buries them. Re-sweeping 850-856 with the current antenna shows real activity at 852.9 / 854.4 / 855.0 / 858.2 MHz (+8 to +11 dB, high variance = voice). The 800 MHz leg is weaker than the 700 MHz control channel but usable.
| TG | Grants | Alpha | Description | Enc | Category |
|---|---|---|---|---|---|
| 17345 | 220 | 17-SGFD Ops | St. George Fire OPS | clear | EBR (17) Fire/EMS |
| 17344 | 110 | 17-SGFD Prevent | St. George Fire Prevention | clear | EBR (17) Fire/EMS |
| 17086 | 57 | 17 JAIL SEC1 | Prison Security | full | EBR (17) Sheriff |
| 17165 | 51 | 17-BRPD DSP1 | Baton Rouge Police Dispatch 1 | partial | EBR (17) BR Police |
| 6039 | 38 | LDWF R4-DISP | Wildlife & Fisheries Region 4 Dispatch | clear | LDWF |
| 17139 | 28 | 17-BRFD DSP1 | Baton Rouge Fire Dispatch 1 | clear | EBR (17) BR Fire |
| 17338 | 22 | — | not in reference DB | ? | — |
| 6848 | 14 | AASI - ZONE 4 | Acadian EMS — EBR/WBR/Iberville/Livingston | clear | EMS Agencies |
| 6846 | 6 | AASI - ZONE 2 | Acadian EMS — Vermillion/Lafayette/St. | clear | EMS Agencies |
Two observed talkgroups (17086 Prison Security, 17165 BRPD Dispatch 1) are flagged encrypted in the database — control metadata only was captured; no audio was decoded and no decryption was attempted.
Local reference DB: 4163 LWIN talkgroups (856 fully encrypted, 114 partial, 3193 clear)
across 243 categories, in reference/lwin_talkgroups.json.
./scripts/lwin_decode.sh 40 # capture 40 s + decode
python3 scripts/annotate_lwin.py # resolve talkgroup IDs to namesscripts/scan_p25band.py captures the whole 768–776 MHz band at 8 Msps and tests
every narrowband carrier for its symbol rate. This distinguishes:
- 4800 baud -> P25 Phase 1 C4FM. The control channel is the one whose 4800 Hz clock is both strongest and most continuous.
The clock estimate is reliable only on strong carriers. On weak or idle channels it returns
junk — see OBSERVATIONS.md §7.1, where a spurious 6000 Hz reading led me to a wrong
conclusion.
Measured here:
| Freq | SNR | Clock | Meaning |
|---|---|---|---|
| 773.05688 | 28.3 dB | 4800 @ 41.4 dB | control channel |
| 769.19434 | 35.7 dB | 4800 @ 19.3 dB | Site 50 (South Baton Rouge) control channel |
| 772.08081 | 23.9 dB | 6000 (unreliable) | Site 50 voice channel — not Site 13 |
| 774.14282 | 30.0 dB | ~6200 (unreliable) | not on Site 13's list |
This is what finally located the control channel. Power-and-variance analysis alone was misleading: 772.08 is a strong continuous carrier, but it is a voice channel belonging to a different site (RFSS 1 Site 50, South Baton Rouge) — which is why every early op25 attempt against it failed.
| Freq | PI | Station | RadioText |
|---|---|---|---|
| 88.5 | 0x6CFA | WJFM / SonLife Radio | SonLife Radio Network www.jsm.org |
| 104.1 | 0x47D6 | "The Vibe" | 104.1 The Vibe - Wanna Be Startin' Somethin' - Michael Jackson |
547 and 559 RDS groups in ~50 s each. Decoded on RTL-SDR #1 (dongle 0 got zero — its antenna is measurably worse).
| Device | ID | Bus | Notes |
|---|---|---|---|
| HackRF Pro r1.2 fw 2026.01.3 | 977c64de2d717413 | 001 | best UHF antenna; used for all P25 work |
| RTL-SDR #0 | SN 00000202 | 003 | PatriotWaves "Skyfall Trunker"; TCXO, measured -0.3 ppm; stock duck antenna |
| RTL-SDR #1 | SN 00000001 | 003 | same unit, second receiver; best FM antenna; only one that decoded RDS |
- rtl_433 silently ignores
.cs8— logsInput format "Unknown", falls back to cu8. HackRF writes signed 8-bit; the RTL ecosystem expects unsigned. Zero decodes, no error. - Over-amplification looks exactly like an empty band.
-a 1on the HackRF with strong local FM raised the whole 24–1800 MHz floor 15–20 dB (pure intermod). Later,-g 62cut the LWIN control-channel SNR from 23.9 dB to 13.1 dB and broke decoding entirely.-l 40 -g 44 -a 0is the working setting here. RTL-SDR #0 needs— WRONG, and harmful. Use-p 25-p 0on both. An early calibration put dongle 0 at +24.6 ppm. Re-measured against the LWIN control channel (773.056250 MHz, a commercial simulcast accurate to a small fraction of a ppm), both dongles read -0.3 ppm — TCXO-grade, consistent with the 28.8 MHz TCXO these units ship with. Applying-p 25therefore introduces ~19 kHz of error at 773 MHz, wider than a whole 12.5 kHz P25 channel, tuning the receiver off the channel. Measured cost ~0.6 dB of SNR on the stock antenna, but with a decent antenna it is the difference between a lock and no lock. Verify withpython3 scripts/cc_snr.py.- op25 vs CMake 4.2: op25 sets
cmake_policy(SET CMP0026 OLD)/CMP0045 OLD, removed in CMake 4.-DCMAKE_POLICY_VERSION_MINIMUMdoes not help. Patch both toNEWand raisecmake_minimum_requiredto 3.10. - HackRF Pro reports an 8–20 Msps floor through gr-osmosdr (HackRF One allows 2).
hackrf_transferstill accepts 2 Msps directly. - op25
rx.pyignores-Fif you also pass-f—elif (self.rtl_found or options.frequency)is evaluated bef 10000 oreelif options.ifile, so it takes the SDR path with a null source and dies. Omit-fwhen decoding a file. - RDS needs >=171 kHz sample rate (57 kHz subcarrier). Lower rates silently drop RDS.
- DVB driver must be unbound: blacklist
dvb_usb_rtl28xxu(done in/etc/modprobe.d/blacklist-rtlsdr.conf). - Live op25 works — but only via SoapySDR, not gr-osmosdr. gr-osmosdr forces an
8 Msps floor on the HackRF Pro (
supported sample rates 8000000-20000000), giving op25 a decim=333 chain that never locked. SoapySDR exposes 1-20 MSps for the same radio; at-S 2000000(decim=83) op25 locks immediately:This was the single blocker on live trunk-following. The file-based pipeline (--args 'soapy=0,driver=hackrf' -N 'AMP:0,LNA:40,VGA:44' -S 2000000 -o 25000scripts/lwin_decode.sh) still works and remains useful for offline re-analysis.
| Band | SNR (HackRF, new antenna) |
|---|---|
| 88–108 MHz FM | +33 dB |
| 162.4 MHz NOAA WX | +10.4 dB |
| 161.975/162.025 AIS | below noise floor |
| 772 MHz LWIN | +17 to +28 dB |
ADS-B (1090), AIS (162) and ISM (433/915) produced no decodes — antenna-limited, not software. A dedicated 1090 MHz ADS-B antenna + LNA would unblock ADS-B; AIS needs a proper marine-band antenna (relevant here given Mississippi River port traffic).
./scripts/lwin_listen.sh # all clear BR-area talkgroups
./scripts/lwin_listen.sh --pd --include-partial 600
./scripts/lwin_listen.sh --tg 17165,17167,17169,17171 --include-partial
./scripts/lwin_listen.sh --pd --list # show selection, do not run
python3 scripts/list_recordings.py| Flag | Selects |
|---|---|
--pd |
police / sheriff dispatch (tag Law Dispatch) |
--pd-all |
police dispatch + talk + tac |
--fire / --fire-all |
fire dispatch (+ tac/talk) |
--ems |
EMS + hospital |
--interop |
interop / emergency ops |
--preset NAME |
pd pd-all fire fire-all ems interop schools publicworks all |
--tag "Law Dispatch,Law Talk" |
select by tag |
--tg 17165,17139 |
explicit talkgroup IDs |
--match REGEX |
regex over alpha / description / category |
--all-areas |
statewide instead of Baton Rouge area |
--list |
print the selection and exit |
Whitelists are generated from the reference DB by scripts/make_whitelist.py.
Recording is clear-only by default.
| Flag | Effect |
|---|---|
--include-partial |
also follow partially-encrypted talkgroups |
--include-encrypted |
also follow fully-encrypted talkgroups (records silence) |
--include-partial is what you want for BRPD and EBR Sheriff dispatch — those are
flagged partial but measured 93% clear (OBSERVATIONS.md §5). op25's -n still silences
any encrypted frames, so nothing unintelligible is written.
Verified on a 200 s --pd --include-partial run: 16 calls, 63.6 s, 16/16 confirmed as
speech, including all 8 calls from partially-encrypted talkgroups (BRPD Dispatch 2 alone
gave 12.1 s). Talkgroups heard: Port Allen PD, BRPD Dispatch 2/4, State Police Troop A
Dispatch 1, EBR Sheriff Dispatch South/Alternate, East Feliciana Sheriff.
Files are named with the talkgroup at save time:
TG17167_17-BRPD-DSP2_20260830-170008.wav Baton Rouge PD Dispatch 2 (12.1s)
TG18000_19-EFSO-DISP_20260830-170203.wav East Feliciana Sheriff (15.1s)
TG5000_SP-A-DISP1_20260830-170051.wav State Police Troop A Disp 1
sdr.db carries full metadata per call — talkgroup, alpha, description, category,
encryption, start time, duration, transcript, and the P25 fields below. It is written
as each call is flushed, so a crash loses at most the call in progress.
The UDP audio stream carries only PCM — no talkgroup ID. udp_audio_record.py
therefore tails op25's log in parallel, tracks the currently-active talkgroup
(voice update: tg(N) / hold active tg(N) / set tgid=N), and stamps it on the file
when the call is flushed. A talkgroup older than 12 s is treated as stale, and a call
whose grant lands just after the first audio packet is back-filled.
Verified on a 160 s run: 10/10 calls labelled, 0 mismatches against an independent re-derivation from op25's own timestamps, and 0 encrypted talkgroups recorded.
Only unencrypted talkgroups are recorded. Two safeguards:
lwin_clear_whitelist.txt(534 clear talkgroups) referenced fromlwin_record.tsv-n(--nocrypt) silences any encrypted traffic that slips through
Encrypted talkgroups observed on this site and deliberately excluded: 17086 Prison Security (full), 17165 BRPD Dispatch 1 (partial), 17133 Baker PD HQ (partial), 17050 Sheriff Dispatch North (partial). No decryption was attempted.
- Run op25 from its
apps/directory.rx.pydoessys.path.append('tdma')— a relative path. Running elsewhere fails withNo module named 'lfsr'; setting PYTHONPATH instead shifts the failure toop25_c4fm_mod(inapps/tx/). - op25 uses the GNU Radio 3.8
wavfile_sinkAPI.p25_decoder.py:117calledwavfile_sink(filename, n_channels, sample_rate, bits_per_sample); GR 3.10 replaced the trailing int with two enums. Patched towavfile_sink(..., blocks.FORMAT_WAV, blocks.FORMAT_PCM_16, False). - Do not pass
-2. LWIN is P25 Phase I;-2setsnum_ambe=2(TDMA) and breaks Phase 1 IMBE voice. Voice updates loggingslot(-)confirm FDMA. - Run op25 under a pty (
script -q -f), notpython3 -u. The log must be written in real time or the recorder mislabels calls, but-ucrashes op25 on Python 3.14 ('_io.FileIO' object has no attribute 'detach'->lost sys.stderr) because op25 reconfigures stdout. A pty makes Python line-buffer naturally.
Control channel is 700 MHz (773.05625) but voice is granted onto 800 MHz (851-853 MHz). A single receiver retunes between the two, so both must be receivable.
These are unencrypted public-safety transmissions — legal to receive in the US and publicly streamed by services like Broadcastify. Recordings nonetheless contain real incident traffic involving real people. They are kept local, are gitignored, and are not redistributed here.
A single radio can either stay on the control channel (see every call, hear none) or follow calls (hear audio, miss grants that occur while it is away). Running both modes gives the full picture.
./scripts/lwin_cdr_run.sh 360 # control-channel only, never retunes
python3 scripts/lwin_cdr.py # parse into a call-detail recordThe trick: point the whitelist at a non-existent talkgroup (lwin_nofollow.txt
contains 999999), so op25 never tunes away and logs 100% of grants.
- 3765 grant events, 14 172 TSBK messages, 20 distinct opcodes
- 101 distinct source radios (unit IDs)
- 19 voice channels in use (851-860 MHz)
- 2479 grants on clear talkgroups, 1167 on encrypted
| Agency | Talkgroups observed |
|---|---|
| LA State Police | Troop A / G / I Dispatch |
| DOTD Motorist Assistance Patrol | Regions 1, 2, 5, 7 |
| Sheriffs | Pointe Coupee, East Feliciana, Iberville, Livingston, St. Bernard |
| Municipal | Port Allen PD, Central PD, Baton Rouge PD (1-4), Baton Rouge Fire |
| Campus / schools | LSU PD, LSU Golf Course, EBR School Board Safety |
| EMS / wildlife | Acadian EMS Zone 4, EBR Medical Common, LDWF Region 7 |
Busiest single talkgroup was 17050 (EBR Sheriff Dispatch North, 494 grants) — which is partially encrypted, so it is excluded from audio recording.
lwin_clear_whitelist.txt was expanded from the 8 talkgroups initially observed to
534 clear talkgroups across every Baton Rouge-area category in the reference DB
(EBR, WBR, Livingston, Ascension, Iberville, E/W Feliciana, Pointe Coupee, LSU, Southern,
State Police Troop A, EMS agencies, LDWF). Zero encrypted talkgroups are in that list.
Ideal setup: park an RTL-SDR on the control channel while the HackRF follows voice.
Measured control-channel SNR on both RTL-SDRs: +2.5 dB (reproduced with
scripts/cc_snr.py) against the ~15 dB P25 needs.
But that measures the stock antenna, not the receiver. Their antennas were never swapped — only the HackRF's was — and on the same channel the HackRF with its swapped antenna reads +28.3 dB. A 26 dB gap is not a front-end difference: the R820T2's noise figure is around 3.5 dB against the HackRF's ~8, so on raw sensitivity these dongles should be comparable or better. The gap is the antenna, which is the one variable never changed — the same trap section 1 already fell into once, when an 850-870 MHz sweep found "nothing" partly because the antenna had not yet been swapped.
Everything except the antenna has now been eliminated as the cause:
| tested | result |
|---|---|
| clock error | -0.3 ppm on both, TCXO-grade. The old "+24.6 ppm, use -p 25" was wrong and was injecting 19 kHz of error |
| receiver health | 33.3 dB SNR on 104.1 MHz FM — the front end is fine |
[R82XX] PLL not locked! |
cosmetic; it fires at 104 MHz too, where SNR is 33 dB |
| tuner IF bandwidth | no measurable benefit — +0.3 dB at 300 kHz, inside run-to-run noise. --tuner-bw exists in op25 now but does not help here |
| tuner gain | best ~4.7 dB at 44.5 dB gain; the front end is not overloading, and turning gain down makes it worse. The long-standing 40 dB was already about right |
| real decode attempt | 0 TSBK with the tuner both wide and narrow, parked on the control channel for 60 s each |
Measured properly the control channel reads ~4.7 dB here, about 10 dB short of the ~15 dB P25 needs.
Two earlier figures in this file were wrong and are retracted. A gain sweep appearing to show +17 dB for turning gain down, and +1.6 dB for clamping the tuner, were both artifacts of measuring at DC: the RTL2832U has a large DC offset spike, and tuning directly to the channel puts the wanted carrier on top of it. Both tools now tune 25 kHz low so the carrier is clear of DC — which is exactly why op25's own working config passes
-o 25000. What has not been tried is the one variable section 7 named and never changed — the antenna. The HackRF reads +28.3 dB on this same channel, and a 26 dB gap between an R820T2 and a HackRF is not a front-end difference.
Know this before trusting any UHF measurement from these dongles. The 28.8 MHz clock produces strong internal spurs at every multiple AND half-multiple:
| multiple | frequency | excess over floor |
|---|---|---|
| 13.5 | 388.800 MHz | +14.2 dB |
| 14.0 | 403.200 MHz | +23.6 dB |
| 14.5 | 417.600 MHz | +33.0 dB |
| 15.0 | 432.000 MHz | +41.0 dB |
| 15.5 | 446.400 MHz | +41.0 dB |
| 16.0 | 460.800 MHz | +43.4 dB |
Off-comb controls at 410.0, 425.0 and 440.0 MHz read +11.2, +1.1 and +10.1 dB.
These pass section 8's retune discriminator — they hold their absolute frequency as the LO moves — because they are fixed internal spurs, not LO-relative artifacts. That test separates LO images from real signals; it does not separate internal spurs from real signals. The comb spacing is what does.
A federal-band survey here initially "found" carriers at 417.599 and 460.801 MHz at +20 and +31 dB. Both are comb entries. Nothing real was found in 406-420 or 450-470.
This rests on geometry, not on a band sweep. An earlier version of this section cited "+24.2 dB at 460-470, its best band" as evidence — that was the 460.8 MHz spur. Re-measured in spur-free windows the picture is much flatter:
| band | excess over floor |
|---|---|
| FM 88-108 | +18.4 dB |
| UHF 465-473 (clean) | +12.0 dB |
| LWIN 800 voice | +9.0 dB |
| UHF 421-429 (clean) | +8.2 dB |
| LWIN 700 control | +6.0 dB |
| VHF 155-162 | +4.3 dB |
UHF still beats LWIN 700, but by 6 dB rather than 18, and LWIN 800 beats one of the UHF windows. Note too that excess-over-floor cannot measure antenna response across bands at all: FM broadcast is 100 kW a few miles away while LWIN is a distant simulcast, so the table mostly reflects transmitter power and distance.
The length argument stands on its own, independent of any of the above:
| frequency | quarter wave | |
|---|---|---|
| supplied 8" duck | 375 MHz | 20.0 cm = 7.87 in |
| LWIN control | 773.06 MHz | 9.70 cm = 3.82 in |
| LWIN voice | 852.5 MHz | 8.80 cm = 3.46 in |
| UHF business | 465 MHz | 16.13 cm = 6.35 in |
An 8-inch whip is a quarter wave at 375 MHz, which is exactly why it peaks at 460. At 773 MHz it is roughly a HALF wave, and a half-wave whip with no ground plane presents a very high feedpoint impedance — a bad match, and most of the signal is lost in it rather than reaching the receiver.
If the duck telescopes — the vendor listing calls it "8" (20 cm) adjustable"
— collapse it to about 3.8 inches and re-measure with scripts/cc_snr.py.
That is a quarter wave at the control channel instead of an awkward half wave,
and it costs nothing. One setting near 3.6-3.8 in covers the 800 MHz voice side
reasonably too.
Note this was measured with the antenna already vertical and window-mounted, so placement is not the remaining variable — length is.
The Skyfall unit has one BNC input, split internally to both RTL receivers. There is no per-receiver antenna connector, and the only antenna it has is the supplied 8" duck. So "swap the antenna onto one dongle" is not a thing that can be done — both receivers always see the same antenna, minus the splitter.
That split is itself a loss nobody has accounted for: a passive 2-way costs 3.5 dB if reactive, 6 dB if resistive, and a unit at this price is as likely to be the latter. Each RTL receiver is therefore working 3.5-6 dB below whatever the antenna actually delivers, on top of the duck being a duck.
Which makes the arithmetic more encouraging than it looks. The HackRF reads
+28.3 dB on this channel with its own antenna. Feed that same antenna to the
Skyfall and each receiver should see roughly +22 to +25 dB after the
internal split — comfortably past the ~15 dB P25 needs, with the --tuner-bw
clamp worth another 1.6 on top.
The decisive test (needs only an adapter, and takes the HackRF offline for a minute): move the HackRF's antenna to the Skyfall's BNC, then
python3 scripts/cc_snr.py -d 1
cd src/op25/op25/gr-op25_repeater/apps
python3 rx.py --args rtl=1 -N LNA:40 -S 1024000 --tuner-bw 300000 \
-T ~/rtl/lwin_cdr.tsv -V -v 2TSBK above zero means the dedicated control-channel receiver works and this section should be rewritten — the prize is measured directly above: 33 talkgroups seen on the control channel against 9 while following voice.
For permanent dual operation the cheapest fix is not a better dongle, it is a second UHF antenna so the HackRF keeps its own. At 773 MHz a quarter wave is 9.7 cm; a quarter-wave ground plane — one 9.7 cm vertical element and four 9.7 cm radials sloping down at 45 degrees, on any panel-mount connector — costs almost nothing and will beat an 8" duck substantially, because the duck has no ground plane to work against. Failing that, an external splitter feeding both radios from the good antenna costs another 3.5 dB on the HackRF path, which it can afford at +28 dB.
(Both dongles are well calibrated: re-measured, each lands within 0.2 kHz of
nominal on the control channel, -0.3 ppm. Use -p 0 for both.)
scripts/find_control.py compares every bin to one global median. Across an 830 MHz
sweep that is wrong — the noise floor is not flat, so whole regions with a locally-high
floor (e.g. 130-137 MHz airband) appeared as ~40 consecutive "signals" at +13 dB.
scripts/find_signals.py replaces it with a rolling-median baseline (default 2 MHz
window) plus a local-MAD z-score. Same 130-960 MHz sweep: 8 real carriers instead of dozens
of phantoms.
| MHz | excess | z | std | Identification |
|---|---|---|---|---|
| 769.1667 | 17.4 dB | 18.3 | 1.22 | LWIN control channel (site 50 / 72) |
| 319.9510 | 13.4 dB | 33.5 | 1.76 | continuous carrier, UHF 225-400 band |
| 157.6961 | 12.9 dB | 26.3 | 5.39 | VHF marine / business, bursty |
| 858.2843 | 10.2 dB | 17.6 | 2.67 | LWIN 800 MHz voice |
| 280.0000 | 10.3 dB | 49.0 | 2.39 | local RFI — see below |
A 38 dB, 0.5 kHz wide, unmodulated constant-envelope carrier at an exactly round frequency. Retuning test (the standard discriminator — a real signal keeps its absolute frequency, an LO artifact keeps its offset):
| LO center | peak offset | absolute |
|---|---|---|
| 279.850 MHz | +150.0 kHz | 280.00002 MHz |
| 280.400 MHz | −400.0 kHz | 279.99998 MHz |
Absolute frequency is stable, so it is not an SDR/LO artifact — it is a real emitter. But sub-kHz bandwidth, no modulation, and a dead-round frequency mark it as a clock or oscillator harmonic from nearby electronics. Worth tracking down if it ever lands on a band of interest.
acarsdec on 130.025 + 131.550 and multimon-ng -a AFSK1200 on 144.390 both ran 7 min and
decoded nothing. Measured band energy on the RTL antennas: +0.9 dB at 131.550 and
+0.3 dB at 144.390 — no usable signal. Same root cause as ADS-B and AIS.
Two acarsdec gotchas found on the way:
- All frequencies must fit one tuner window (~2 MHz).
129.125 … 131.725is rejected withFrequencies too far apart. -r <device>must come last, immediately before the frequency list. Putting-g 49.6after-rmakes acarsdec parse the gain as a frequency:Invalid frequency 49600000.
Everything still blocked (ADS-B, AIS, ACARS, APRS, a dedicated LWIN control-channel receiver) is blocked by the RTL-SDR antennas, which were never swapped. The single cheapest unlock is a coax splitter feeding the good antenna to an RTL-SDR, or a second wideband antenna. That one change enables a dedicated control-channel receiver, which would recover the ~60% of calls the single radio misses while it is away following a call.
Measured, not assumed. The P25 ESS (Encryption Sync Sequence) in the LDU2 voice frame
carries ALGID, KID and the message indicator in the clear, so the cipher can be
identified without any decryption. op25 prints it at -v 10 (p25p1_fdma.cc:279).
# follow the encrypted talkgroups (no -n, no audio recorder) and read ESS headers
script -q -f -c "cd <op25>/apps && python3 rx.py ... -T lwin_enc.tsv -V -v 10" results/enc_id.log
python3 scripts/identify_encryption.py| ALGID | KID | Count | Algorithm |
|---|---|---|---|
0x80 |
0x0 | 138 | CLEAR — unencrypted |
0xAA |
0x8 | 9 | ADP / RC4 (Motorola "Advanced Digital Privacy", 40-bit) |
0xC4 |
0x9726 | 1 | not a valid P25 ALGID — bit error (marginal 800 MHz SNR) |
| TG | RR flag | Alpha | Observed |
|---|---|---|---|
| 17050 | partial | 17-SO DISP N | clear ×39 (+1 corrupt) |
| 17171 | partial | 17-BRPD DSP4 | clear ×30, ADP ×2 |
| 17165 | partial | 17-BRPD DSP1 | clear ×18, ADP ×6 |
| 17086 | full | 17 JAIL SEC1 | clear ×23 |
| 17051 | partial | 17-SO DISP S | clear ×21 |
| 17133 | partial | 17-BAKER PD HQ | clear ×7 |
| 17169 | partial | 17-BRPD DSP3 | ADP ×1 |
The only encryption algorithm observed is ADP (ALGID 0xAA), key ID 0x8, and only on Baton Rouge PD dispatch talkgroups (17165 / 17169 / 17171) — all sharing one key ID. No AES-256 (0x84), AES-128 (0x85) or DES (0x81/0x83) was seen on the sampled talkgroups.
Two things worth noting:
- The "partial" flag in RadioReference is accurate: 93% of frames on these talkgroups
were
0x80clear. They encrypt selectively, not continuously. - 17086 "Prison Security" is flagged
fullin RadioReference but transmitted0x80clear in all 23 observations here — the database flag looks stale, or those particular transmissions were not encrypted.
Caveats: ~4 minutes, only the 9 whitelisted talkgroups, and the 800 MHz voice channels are
marginal here (the stray 0xC4 is direct evidence of frame corruption). Other agencies on
LWIN (State Police, DOTD, other parishes) were not sampled and may differ. A longer run
across more talkgroups would firm this up.
No decryption was attempted and no key material was recovered — only the cleartext algorithm identifier was read. Encrypted talkgroups remain excluded from recording.
docs/2026-08-31-session-log.md — what was built, what was measured, what was
ruled out, and where to resume. Includes the RF investigation into why the
RTL-SDRs cannot decode the control channel, which ends at a testable hypothesis
(the supplied antenna is resonant at 375 MHz, not 773) pending a new antenna.
Everything lives in sdr.db (SQLite): 4,163 talkgroups, 149 sites, 243
categories, every recorded call with its transcript, and the control-channel
grant log. The JSON files under reference/ and the .wav/.txt files under
recordings/ remain on disk — the database indexes them, it does not contain
them — and op25 still reads its whitelist and talkgroup TSV as files.
python3 scripts/import_to_sqlite.py # (re)build from the flat files
python3 scripts/import_grants.py results/lwin_cdr.log
python3 scripts/list_recordings.py -n 20 --search dispatchThe database is gitignored: it holds transcripts of live public-safety traffic
and per-radio identifiers, the same posture as recordings/. It is fully
rebuildable from the files on disk.
udp_audio_record.py used to write recordings/calls.json in its finally
block containing only that session's calls — a truncating write, so every
run discarded the metadata for every recording before it. A 60-second session
took the file from 2,953 entries to 7; the next one took it to 1. Transcripts
merged in by stt_watch.py were clobbered a few minutes later by the same
write, which is why no transcript ever survived there.
Calls are now INSERTed as they are flushed. Nothing was lost — every field is
derivable, so scripts/import_to_sqlite.py reconstructed the full history from
the filenames, the WAV headers and the reference DB.
Beyond the talkgroup, each call records what op25 observed for it:
| Field | Source | Note |
|---|---|---|
freq |
voice update: tg(N), freq(H) |
the voice channel that call used |
algid / keyid / mi |
ESS header in the LDU2 frame | needs --ess (op25 -v 10) |
nac / sysid |
NID / rfss_sts_bcst |
0x1bd on this system |
src_addr |
grant TSBK srcaddr |
the transmitting radio |
rfss / site |
rfss_sts_bcst |
control-channel only, see below |
ESS is the authoritative encryption signal for a call, and it disagrees with
the reference DB's static per-talkgroup flag in both directions — TG 17086 is
flagged full upstream but transmitted algid 0x80 (clear) in all 23
observations, and TG 17165 is flagged partial while most of its calls are
clear. The console shows both: Enc is the talkgroup's label, Observed is
what this call actually transmitted. --ess costs roughly 10x the log volume
(~15 MB/hour against ~2.4), so it is off by default.
rfss/site are normally NULL on recorded calls, and that is structural.
They come from rfss_sts_bcst, a control-channel broadcast: 891 occurrences in
a control-channel-only capture, zero in a voice-following log. Capturing them
per call would need the control channel watched while voice is followed —
two receivers — and both RTL-SDRs measure +2.7 dB on the control channel
against the ~15 dB P25 needs (section 7). The grants table carries the site
for control-channel captures instead.
A note on terminology, since it is easy to get wrong: DUID is not the transmitting radio.
p25_framer.cc:101-102extracts two fields from the NID —nac = (acc >> 52) & 0xfff(12 bits) andduid = (acc >> 48) & 0x00f(4 bits). DUID is a 4-bit frame type (HDU0x0, LDU10x5, LDU20xa, TDU0x3, TSBK0x7, TDULC0xf), which is why op25 testsduid == 0x3 || duid == 0xfto detect voice termination. The 24-bit "who transmitted" field is the Source ID, logged assrcaddrand stored assrc_addr. It is sparse: 3,223 of 3,765 real grants carrysrcaddr=0.
Run all of these FROM THE REPO ROOT.
npm test # both suites
./node_modules/.bin/vitest run # 317 TypeScript
python3 -m unittest discover -s scripts/tests # 455 Pythoncd scripts/tests && python3 -m unittest discover -s . looks equivalent and is
not. scripts/tests/test_capture_control.py imports
from scripts.capture_control import ... — a package path that only resolves
with the repo root on sys.path. From inside the directory it fails with
ModuleNotFoundError: No module named 'scripts', and unittest reports it as
one collection ERROR among otherwise passing tests:
Ran 381 tests ... FAILED (errors=1) <- wrong invocation
Ran 455 tests ... OK <- from the repo root
That failure is manufactured by the invocation, and it is convincing: it
names a real test file, it is stable across runs, and everything else passes.
It was mistaken for a pre-existing repo defect for a whole session and written
into two commit messages as "the lone discovery-path error predates this work".
It does not exist. npm test has always exited 0.
Two lessons worth keeping, since this repo's failures are usually of exactly this shape: check how a project runs its own tests before explaining why they fail, and treat a diagnostic's output as evidence about the diagnostic until it has been shown otherwise.
The TypeScript tests run against the real sdr.db rather than fixtures,
deliberately: every data bug here has been a disagreement between assumed and
real data, and a fixture reproduces the assumption rather than the data.
A Nuxt 3 + PrimeVue 4 app (pages/, components/, server/api/) replaces the
old Python stdlib server.
pnpm install
./node_modules/.bin/nuxt dev --host 0.0.0.0 # http://0.0.0.0:3000, hot reload
# or, for production
./node_modules/.bin/nuxt build && node .output/server/index.mjsCommands go through
./node_modules/.bin/rather thanpnpm run: pnpm 11.17 aborts every script run withERR_PNPM_IGNORED_BUILDSfor esbuild, whose platform binaries are in fact installed.pnpm installitself is fine.Reaching the dev server by hostname needs that host in
NUXT_ALLOWED_HOSTS(see.env.example) — Vite refuses unrecognised Host headers as a DNS-rebinding guard, and bare IPs are exempt but names are not. This is dev-only: Nitro's production server performs no Host check, so the variable is inert undernode .output/server/index.mjs. It is a convenience, not a security control — protection against a hostile page driving the console is the same-origin guard inserver/utils/guards.ts, which applies to both.
Open http://10.56.1.77:3000/ (or http://127.0.0.1:3000/) — three panels:
- Listen & Record — select by preset, explicit talkgroup IDs, tag
(
--tag "Law Dispatch") or regex (--match BRPD); statewide via--all-areas; independent--include-partial/--include-encryptedswitches; Whisper STT; duration. Start spawnsscripts/lwin_listen.shin its own process group and captures its output toweb/listen.log; Stop sends SIGINT to the group, whichlwin_listen.sh's owncleanuptrap turns into an orderly teardown of op25, the recorder and the STT watcher. Status polls every 5 s and shows the live call count, the pid, and what the session is following. - Recordings — every
recordings/TG*.wavwith its metadata and Whisper transcript fromsdr.db. Search covers talkgroup, alpha, description, category, filename and transcript text; the encryption filter offers clear / partial / full / unlabelled.[BLANK_AUDIO]transcripts (a silenced encrypted burst, not speech) are dimmed. Audio is served with HTTP Range support — including suffix ranges — so seeking works. - Talkgroups — the reference DB, Baton Rouge area (601) or statewide (4163),
filterable by category and encryption, searchable including the
tagfield; rows in the currentlwin_active_whitelist.txtare marked and highlighted. Note that file is regenerated by every recording run, so its size tracks the last preset used.
A running session is tracked in web/listen.pid, so restarting the web server does not
lose the ability to stop it. All four web/listen.* sidecar files are gitignored.
pages/index.vue the Strip Bay — the only page
components/bay/ CommStack · CallStrip
composables/ useScannerFeed · useArchive
server/api/ listen/ · recordings/ · talkgroups/ · config/ · transcribe/
server/utils/ paths · files · talkgroups · processes · session · recordings
assets/css/compat.css bridges PrimeFlex 3 theme classes onto PrimeVue 4 Aura tokens
Unit tests (./node_modules/.bin/vitest run) cover the server utils, including
assertions against the live reference DB (4163 talkgroups, 601 in the Baton Rouge area).
sdr.db SQLite: talkgroups, sites, calls, transcripts, grants [gitignored]
pages/ components/ server/ <- Nuxt web console (port 3000), see above
web/ listen.* runtime state <- pid/log/config, gitignored
scripts/ lwin_listen.sh <- start listening (one command)
sdr_db.py <- schema + the writers the recorder uses
import_to_sqlite.py <- (re)build sdr.db from the flat files
import_grants.py <- control-channel grants -> the CDR table
list_recordings.py <- show what was recorded (reads sdr.db)
lwin_cdr_run.sh / lwin_cdr.py <- full call-detail record
find_signals.py <- wideband survey (rolling baseline)
udp_audio_record.py lwin_decode.sh annotate_lwin.py
scan_p25band.py sweep_peaks.py find_control.py fetch_lwin_db.py
stt_transcribe.py <- batch-transcribe existing recordings (whisper.cpp)
stt_watch.py <- live STT watcher (transcribes new .wav as they land)
tests/ <- 35 unittest cases over the above
results/ sweeps, op25 logs, RDS json, SPECTRUM_REPORT.md
reference/ lwin_talkgroups.json lwin_sites.json lwin_categories.json
(import sources for sdr.db; op25 still reads its whitelist as a file)
captures/ raw IQ (.cs8) and op25 input (.cfile) [large, gitignored]
src/ op25 (patched), rtl-ais, redsea, acarsdec
tools/ whisper.cpp (built binaries + shared libs)
models/ ggml-small.en.bin (466 MB)
Audio from udp_audio_record.py is 8 kHz mono S16LE PCM. Transcription is local via
whisper.cpp (CPU-only; the host's GPU is busy with vLLM), model small.en.
Batch:
python3 scripts/stt_transcribe.py --force # transcribe all existing .wavWrites recordings/TG*.txt (one per wav) and indexes them in sdr.db, where FTS5
makes them searchable from the console. The .txt beside the .wav stays the
durable copy; the database only indexes it.
Live (alongside lwin_listen.sh):
./scripts/lwin_listen.sh --stt --pd --include-partial 600Launches scripts/stt_watch.py in the background. Every new TG*.wav that
udp_audio_record.py saves is picked up within a few seconds and transcribed with the same
small.en model, then indexed in sdr.db. Ctrl-C stops op25, the recorder, and the
STT watcher together.
Performance on this box (32 cores): ~1.3 s per 3 s clip, ~103 s for 82 clips total. No API key, no per-call billing — fits the "recordings stay local" privacy posture (§9).
The model file (models/ggml-small.en.bin) and the built whisper.cpp binaries
(tools/whisper.cpp/build/bin/whisper-cli) are the only new artifacts. To use a different
model, pass --model <path> to either script.
LWIN's packet data is readable (see patches/README.md and scripts/p25_packet.py),
but only the outbound half. iden_up id 1 carries toff +30 MHz and id 0 carries
-45, so every grant assigns a channel pair:
| leg | downlink | uplink |
|---|---|---|
| 700 | 769–775 | 799–805 |
| 800 | 851–860 | 806–815 |
Our receivers are on the downlink, so we see the system asking radios for their location and never the answers. LRRP position reports travel inbound.
Bandwidth is not the obstacle. The two uplink bands are adjacent — 799–815, 16 MHz — and one HackRF at 20 Msps covers both (±8.50 MHz around 807).
Signal is the obstacle, and it was measured rather than assumed. Two hours on 803.05625 — the control channel uplink, where every radio in range transmits registrations, affiliations and service requests — on the better of the two antennas:
elapsed 01:59:30 log lines 6918 frames 0 candidate frame syncs 0
The second zero is the informative one. On the downlink the framer regularly logs
p25_framer::load_nid() error check failed — it found something frame-sync-shaped and
rejected it on BCH. Here it found nothing to reject in two hours. That distinguishes
"too weak to decode" from "not present", and it is the latter.
Positive control, same command and software, pointed at the CC downlink for 60 s:
624 frames — 496 TSBK, 48 LDU1, 31 TDU15, 1 HDU
A 60-second version of the uplink test came first and was NOT conclusive: uplink carries traffic only when an in-range radio transmits, so a short empty window proves nothing. Two hours is what makes the negative worth recording.
Why, roughly. A site runs ~100 W per channel from a tower with line of sight over the area; a portable is ~3 W (−15 dB) and a mobile ~25 W (−6 dB), at street level, often inside a vehicle or building — and our antenna is window-mounted, in the clutter too. Uplink reception is a proximity problem: you hear the subscribers near you, not the ones near the tower.
So no purchase fixes this at this location. A 790–820 MHz band-pass filter, an LNA, or a 12-bit receiver all rescue marginal signal, and there is nothing marginal here — a sweep did show strong cellular uplink at 827–829 MHz (+24.4 dB, above anything of ours, and worth filtering if this is ever retried), but filtering cannot create a frame sync where there are none. What would change the answer is siting: an antenna near where radios actually operate — a station, a garage, a district — or a directional antenna aimed at one such concentration. Both are physical decisions, not equipment.
Even full success would yield the radios near that antenna, not the fleet.
Everything here is receive-only of unencrypted, publicly broadcast signals, which is legal in the US. Encrypted talkgroups were not decrypted (several EBR law-enforcement talkgroups are flagged full/partial encryption in the reference DB and were left alone). Cellular bands were not touched (18 USC 2511). No digital-pager message content was logged.