nocturne-fieldrecordist

Combined BirdNET + Nocturne bioacoustic classification for wildlife field-recordist libraries.

• BirdNET ≈ 3,000 bird species (runs locally via birdnetlib)
• Nocturne ≈ 2,182 non-bird species — insects, amphibians, non-bird mammals, reptiles (Stratus Labs hosted API at nocturne.runstratus.com)
• Scan a directory → get a species-tagged SQLite catalog searchable by species, model, confidence
• Ships with a Reaper action that adds species markers to a selected media item

Built by Stratus Labs. Nocturne itself is Apache-2.0; BirdNET is licensed separately (see the BirdNET-Analyzer repo).

Install

pip install nocturne-fieldrecordist

This pulls in nocturne-client, birdnetlib, and tensorflow (BirdNET's runtime). Total install ~2 GB. First run of BirdNET downloads the model (~50 MB, cached).

Scan a library

nocturne-fr scan ~/recordings --db library.sqlite

Recurses into subdirectories; audio extensions: .wav .flac .mp3 .m4a .ogg .aif .aiff. Skips files already in the DB unless you pass --reprocess. Progress prints to stderr; a JSON summary lands on stdout.

Classify a single clip

nocturne-fr classify path/to/clip.wav

Prints a JSON list of detections from both models.

Query the catalog

# All Nocturne detections above 0.3 confidence, limit 50
nocturne-fr query --db library.sqlite --model nocturne --min-conf 0.3 --limit 50

# Everything matching "wren" (either BirdNET or Nocturne, any confidence)
nocturne-fr query --db library.sqlite --species wren

Python API

from nocturne_fieldrecordist import DualClassifier, scan_path

dc = DualClassifier(birdnet_min_conf=0.5, nocturne_min_conf=0.2, nocturne_top_k=5)
for det in dc.classify("recording.wav"):
    print(det.source_model, det.species, det.confidence, det.start_time_s)

# Or scan a whole tree:
summary = scan_path("~/recordings", db_path="library.sqlite", dc=dc)

Reaper integration

Install the action from nocturne_fieldrecordist/reaper/nocturne_reaper.py:

1. Set Reaper's ReaScript Python interpreter to the same env where you pip installed this package (Preferences → Plug-ins → ReaScript).
2. Actions → Load ReaScript → select nocturne_reaper.py.
3. Optional: bind a keyboard shortcut.

Select a media item, trigger the action, and get one Reaper marker per detection: [bir] Passer domesticus (0.87) or [noc] Neoconocephalus triops (0.42). BirdNET markers get one color, Nocturne another.

Tunables

• --birdnet-min-conf (default 0.5) — BirdNET's author recommends ≥0.5.
• --nocturne-min-conf (default 0.15) — Nocturne v1's per-class-calibrated thresholds sit around 0.15-0.30 across the vocab.
• --top-k (default 3) — how many Nocturne predictions to keep per file.

Nocturne API auth

Anonymous is fine (rate-limited to 10 req/min per IP). For batch libraries pass a paid-tier key:

export NOCTURNE_KEY=ck_...
nocturne-fr scan ~/recordings --db library.sqlite --nocturne-key $NOCTURNE_KEY

Sign up at nocturne.runstratus.com/pricing.

SQLite schema

Two tables: recordings (one row per audio file) and detections (one row per model prediction). See schema.sql for indexes + full columns. The catalog file is portable — copy it, ship it to a Flask app, load it in DB Browser, whatever fits your workflow.