Python tool for aircraft Doppler audio geolocation using historical ADS-B flight data (Source-available)
Real audio collection and Doppler-candidate review
All new recordings in this collection were processed with the automatic audio scanner only. No flight matching, aircraft association or geolocation was run on the collected clips. No coordinates, flight IDs or recording windows were scanner inputs. Downloaded media, detailed ridge outputs and review plots stay local.
The broad video survey and the smaller candidate index serve different purposes: the survey preserves positive, negative and failed outcomes; the index collects the strongest reviewed real-recording leads for follow-up. Neither is an accepted aircraft-localization benchmark yet.
Broad online-video survey
The expanded case table and summary contain 314 distinct YouTube video IDs. The final per-ID outcomes are 273 successful scans, 41 unsuccessful acquisitions, 47 successful scans with moving-tone proposals and 226 successful scans with no proposals. All 47 proposal cases received a quick single-reviewer spectrogram screen: five were labeled candidate Doppler-like, five looked plausible but showed possible edits or repetition, and 37 remained ambiguous. These labels do not confirm aircraft identity or continuous recording time.
Successful-scan proposals by query category were:
| Search category | Scanned | With proposals |
|---|---|---|
| Aircraft-targeted | 62 | 18 |
| Outdoor walks | 48 | 10 |
| Site-named aircraft | 26 | 8 |
| Adverse conditions | 39 | 4 |
| Birds and nature | 26 | 3 |
| Indoor | 29 | 3 |
| Phone movement | 43 | 1 |
These categories describe the searches, not verified microphones or recording conditions. Results are query-selected, not random or representative of YouTube. Most scans cover the first 180 seconds, or the entire source if shorter; this is not full-video screening for long sources. The three indoor proposals were all ambiguous. A no-proposal result applies only to the analyzed excerpt.
The original 392-source survey snapshot is retained unchanged as a historical artifact. The expanded table reconciles it with retries and the separate site-named search, deduplicates by video ID and preserves unsuccessful acquisition outcomes and multiple attempts.
Licensed aircraft-audio scanner sample
The AeroSonicDB scanner-only report records a fixed, stratified random scan of 155 real aircraft and ambient clips from a research dataset. Scanner results and limitations are documented separately from the dataset's labels and flight metadata. AeroSonicDB is licensed CC BY-NC 4.0; no media is copied into this repository. Its public coordinates refer to nearby public places rather than exact microphones, so these clips do not establish location truth for geolocation evaluation.
A follow-up review selected 20 known-site candidate clips from that sample. All 20 were publisher-labeled aircraft recordings, emitted at least one scanner proposal, and were visually reviewed as having a clear descending tonal ridge. They are linked in the scanner-facing manifest; publisher aircraft/date/time/site/microphone metadata and the approximate public site markers are in separate event-reference and location-marker tables. The scanner manifest contains no truth labels or coordinates. No flight matching or geolocation was run on these clips.
The local audio hardlinks are at
/home/cinna/.local/state/dopplergeist-benchmark-candidate-audio-20261103/aerosonic-known-site/;
they are not redistributed in this repository. They can be reacquired from the
AeroSonicDB v1.1.2 record under CC BY-NC
4.0. The source labels provide recording date/time strings, microphone IDs and
one of three site IDs, but the timezone is not stated. The three published
coordinates are nearby public-place markers with unquantified microphone offset.
This is a 20-clip known-site scanner candidate collection, not a 20-point
microphone ground-truth benchmark and not evidence of 100 m localization
accuracy.
Publisher-documented aircraft scanner control
The DLR RECORD 139 scanner check uses one center-microphone channel from a real, publisher-identified A320 flyover recording. The automatic scanner found no ridge proposals in this seven-second excerpt. That is a scanner miss on this excerpt, not proof that no Doppler structure exists in the full array recording. The source contains microphone and aircraft geometry only in a relative coordinate frame; the inspected HDF5 does not provide geodetic microphone coordinates or an absolute UTC clock. No flight matching or geolocation was run, and this control does not count toward the 20-known plus five-unknown location benchmark target.
Candidate index
The 25-row candidate index contains:
- 20 Eclipse Soundscapes site recordings with a reviewed descending tonal band and a publisher-reported site coordinate. Two of the 20 have weak or interrupted visual support. The coordinate reference is kept separately in the truth-separated table. The publisher coordinate is for a deployment site; microphone placement error is unknown. Filename-derived times are claims, not independently verified UTC.
- Five YouTube candidates labeled candidate Doppler-like in the visual review. Their exact receiver coordinates are unprovided; some titles mention a coarse place. Aircraft identity, recording UTC and independence from editing are not established.
The separate five-source unknown-exact-point shortlist contains real clips with visually plausible descending patterns and positive scanner proposals. Their exact receiver coordinates and absolute recording times are unverified; one is an edited compilation, and another may contain overlapping harmonic groups from a single pass. This is candidate material, not five validated unknown-location benchmark examples.
The Eclipse and YouTube candidates are real recordings and real moving-tone evidence, but the source of the tonal bands has not been confirmed as aircraft. The site coordinate is not a surveyed microphone coordinate. These clips are useful candidate material for further evidence gathering; they cannot yet measure 100–1,000 m localization error. The requested 20 known-location usable aircraft clips plus five usable unknown-location aircraft clips remains unmet until aircraft/source identity and the needed receiver position/time evidence are independently established.
The candidate exporter uses only the committed scanner/review manifests and local publisher metadata. It creates the location table separately:
python scripts/export_audio_candidate_index.py \
--truth-root /home/cinna/.local/state/dopplergeist-eclipse-corpus-20261103 \
--index benchmarks/audio-candidate-index-20261103.csv \
--truth benchmarks/truth-separated/audio-candidate-location-references-20261103.csv
Never pass the truth-separated coordinate table into the scanner or use it to narrow an evaluation search. The exporter does not download or rescan audio.
Acquisition and scanner reproduction
The broad survey used the predeclared query plan
online-audio-search-plan-20261103.json
and collect_audio_corpus.py. Collection
clips use the first available 180 seconds, or the full shorter source; sample
zero and upload time are not recording UTC. The collector's scanner settings are
recorded in the saved scan JSON under the local corpus roots. No source video or
audio is committed.
The expanded table can be regenerated without download or rescanning when its six local collection roots remain available:
python scripts/export_youtube_survey_20261103.py \
--state-root /home/cinna/.local/state \
--output /tmp/youtube-expanded-scan-20261103.csv
A fixed-seed sample of 200 videos was selected without replacement from the
query-derived YouTube catalog and scanned in scanner-only mode. It is random
within that catalog, not representative of YouTube. Of 200 selected videos, 198
were scanned and two failed download; 24 scanned excerpts produced scanner
proposals, while 174 had no proposal in the first 180 seconds. All 24 proposal
plots were visually reviewed for trace shape only; aircraft identity remains
unverified. No flight matching or geolocation was run. Per-video outcomes,
summary counts and review limits are in the random-sample report.
Audio, detailed scanner artifacts and logs remain local under
/home/cinna/.local/state/dopplergeist-youtube-random-sample-20261103/youtube-only-200/.
Source rights and continued online availability are not established by this survey. Retain failed downloads, ambiguous proposals and no-proposal excerpts in any future update.