BRIARHILL BACKYARD BIRD LOG
Methodology
A backyard acoustic station records the soundscape, identifies likely bird calls locally, and shares a growing archive of detections.
- Recording format
- 48 kHz mono, 16-bit WAV
- Recording block
- 5 minutes followed by analysis
- Detection threshold
- 0.50 BirdNET model score
- Automatic audio retention
- Up to 7 days subject to storage limits
01 / THE STATION
Hardware & location
- Computer
- Raspberry Pi running Debian Linux. Recording and BirdNET analysis run on the station, not in the browser.
- Audio input
- A microphone connected through a UGREEN USB audio adapter. ALSA captures a single audio channel.
- Local storage
- Detection results, short audio excerpts, review marks, and an upload queue are stored on the Pi.
These observations come from sound alone. Camera, weather, and motion sensors are not inputs to the bird counts shown here.
Briarhill
Station locationLoading station location...
02 / FROM SOUND TO RECORD
The recording workflow
- 1
Record
ALSA records a five-minute block. Recording pauses while that block is analyzed, so this is not a gapless recording or a livestream.
- 2
Identify
BirdNET Analyzer assigns species labels and scores to sound segments. The monitor retains detections at a score of 0.50 or higher.
- 3
Keep excerpts
A detection gets up to two seconds of context on each side. A typical three-second detection produces a seven-second clip; excerpts are capped at 15 seconds.
- 4
Synchronize
A scheduled job sends new metadata and eligible audio to the public archive about every five minutes. Pending uploads retry after a failure.
Species validation is a separate owner review step. Automatic publication does not mean the identification has been confirmed.
03 / SOFTWARE
Local analysis, public archive
- Recording & analysis
- Python coordinates ALSA
arecordand BirdNET Analyzer. A regional species filter narrows the candidates, but does not confirm an identification. - Review & scheduling
- A local Python dashboard supports listening and species validation. SQLite keeps the upload queue; systemd runs the recorder, dashboard, and upload schedule.
- Public website
- PHP endpoints on GoDaddy hosting with Plesk read the MySQL archive. HTML, CSS, and JavaScript display detections and hourly/daily statistics.
- Spectrograms
- NumPy calculates a short-time Fourier transform. Images show 0–8 kHz, with higher frequencies at the top and time running left to right. Per-clip contrast reveals faint features without changing the audio.
04 / INTERPRETATION
What the numbers mean
- Detections are sound segments, not individual birds. Repeated or overlapping calls can create multiple records.
- Species totals are distinct predicted labels. Hourly, daily, and whole-period totals are calculated separately; hourly species totals cannot be added to get daily richness.
- A model score is not a calibrated probability. Background noise, overlapping calls, and species-specific errors affect identification.
- Validated means owner-confirmed. Unvalidated is not the same as incorrect. Audio availability is independent of validation.
- Times are approximate station-local times. They are reconstructed from recording timestamps and segment offsets. An empty hour is not evidence that no birds were present.
- Counts are not adjusted for recording effort. Analysis pauses and other gaps are not measured by the statistics. Spectrogram colors are relative to each clip, not comparable sound-pressure measurements.
05 / RETENTION & PRIVACY
Short-lived audio, lasting records
Full recording blocks stay local and are deleted after analysis. Local excerpts expire after seven days, or sooner if their combined size exceeds 1 GiB. Local result files expire after 30 days.
Automatically published clips are eligible for removal seven days after publication, or earlier under the public storage limits of 512 MiB / 2,000 clips. Cleanup runs during media synchronization. Older manually published clips are not automatically expired by this policy.
Audio removal does not delete the detection record, validation, or statistical counts. The public metadata archive is independent of local file retention.
Human sounds & owner review
Recordings with human, person, or speech labels are excluded from automatic publication. This is not reliable speech detection: background conversation or other private sounds can still be published.
The owner can remove audio or flag it as inappropriate. Either action removes its public audio and spectrogram and blocks reupload; it does not change the species validation. Previously downloaded or cached copies cannot be recalled.