Passive acoustic monitoring (PAM) is the practice of deploying autonomous recorders to capture sound continuously over weeks or years — turning biodiversity into a streamable time series. AudioMoth, SongMeter, and Wildlife Acoustics SM4 are the dominant hardware; BirdNET and Perch are the leading deep-learning classifiers; Arbimon, ARBIMON-PAM and Rainforest Connection provide cloud infrastructure. The discipline is moving rapidly from boutique fieldwork to global infrastructure.
Passive Acoustic Monitoring
Passive Acoustic Monitoring (PAM) is the deployment of autonomous recording units (ARUs) to capture soundscapes continuously, without human presence.
| Device | Frequency | Feature | Use |
|---|---|---|---|
| Song Meter SM4 | 20 Hz – 192 kHz | Industry standard | Terrestrial multi-taxa |
| AudioMoth | 8 Hz – 384 kHz | Open-source, £40 | Mass deployment, bats |
| Swift | 20 Hz – 48 kHz | Long autonomy | Remote sites, birds |
| AMAR-G4 | 1 Hz – 180 kHz | Marine hydrophone | Cetaceans, fish |
| HARP | 10 Hz – 100 kHz | Seafloor deployment | Deep-sea, whales |
Machine Learning for Soundscape Analysis
CNN architectures applied to mel-spectrograms: BirdNET (Cornell, 6000+ species), Perch (Google), AVES (self-supervised). Unsupervised methods include NMF:
$\mathbf{V}$ spectrogram, $\mathbf{W}$ spectral basis (acoustic atoms), $\mathbf{H}$ temporal activations. Enables discovery of novel sound types.
The end-to-end pipeline
- Field deployment: choose number, spacing, schedule of recorders to balance area coverage against per-site density.
- Data offload: SD card swap, cellular telemetry, or LoRa-mesh. Bandwidth quickly becomes the bottleneck.
- Pre-processing: noise reduction, segmentation into vocalisation events, normalisation.
- Classification: deep-learning models identify species per snippet with class-conditional posteriors.
- Ecological inference: occupancy models (Royle-Nichols, MacKenzie et al.) translate detections into species presence under imperfect observation.