Noisy Pi

Ambient noise monitoring for Raspberry Pi, designed to run alongside BirdNET-Pi.

Noisy Pi captures audio from BirdNET-Pi's Icecast stream, analyzes noise levels and frequency content, detects anomalies, and provides a web dashboard for visualization.

• Non-interfering: Uses BirdNET-Pi's existing Icecast audio stream
• Noise metrics: Mean, max, and min dB levels
• Percentiles: L10, L50, L90 statistical levels
• 7 Frequency bands: Detailed coverage from 0-24kHz
• Spectral features: Centroid, flatness, dominant frequency
• Silence detection: Percentage of quiet time per sample
• Full spectrogram: 256-bin FFT with 10 snapshots per sample

• Statistical baseline: Learns normal patterns per hour/day-of-week
• Z-score anomalies: Flags measurements that deviate significantly
• Visual indicators: Anomalies highlighted in dashboard
• Optional snippets: Save audio clips of anomalies for review (privacy-aware, opt-in)

The dashboard has multiple tabs with rich visualization:

• Stats overview: Current level, max, min, centroid, silence %, anomaly count
• Sound levels chart: Time-series of mean, max, and L90 levels
• 7-band heatmap: Clickable frequency vs time visualization
• Anomaly chart: Z-score timeline with threshold indicator
• Recent measurements: Table with inline annotation editing
• Audio snippets: Playback and management of captured anomalies

• Full spectrogram view: Detailed band-based visualization
• Colormap selection: Viridis, Plasma, Inferno, Magma
• Measurement detail: Click to view individual sample spectrum
• Detailed metrics: Centroid, flatness, dominant frequency

• Period selection: Today, This Week, This Month, All Time
• Stats cards: Aggregate metrics for selected period
• Hourly pattern: Bar chart of today's activity
• Weekly baseline heatmap: Learned patterns by hour and day

• Date range picker: Select custom date ranges
• Historical charts: Visualize past data
• Data export: Download CSV for external analysis
• Full data table: All 7 frequency bands displayed

• Anomaly threshold: Adjust Z-score sensitivity
• Audio snippets: Enable/disable anomaly recording
• Snippet duration: Configure recording length
• Auto-refresh interval: Set dashboard update frequency

• Raspberry Pi with BirdNET-Pi installed and running
• BirdNET-Pi's Icecast stream enabled (default configuration)
• ffmpeg, PHP, SQLite3, Python3

curl -s https://raw.githubusercontent.com/andjar/noisy_pi/main/install.sh | sudo bash

Or clone and install:

git clone https://github.com/andjar/noisy_pi.git
cd noisy_pi
sudo bash install.sh

After installation, access the dashboard at:

• http://your-pi-hostname.local:8080
• http://your-pi-ip:8080

# Check service status
sudo systemctl status noisy-capture
sudo systemctl status noisy-web

# View logs
journalctl -u noisy-capture -f
tail -f /var/log/noisy-pi/capture.log

# Query database
sqlite3 /var/lib/noisy-pi/noisy.db "SELECT * FROM measurements ORDER BY id DESC LIMIT 10;"

# Restart services
sudo systemctl restart noisy-capture
sudo systemctl restart noisy-web

Edit /opt/noisy-pi/config/noisy.json:

{
    "icecast_url": "http://localhost:8000/stream",
    "sample_rate": 48000,
    "sample_duration": 30,
    "sample_interval": 30,
    "anomaly_threshold": 2.5,
    "baseline_min_samples": 100,
    "snippet_enabled": false,
    "snippet_duration": 5,
    "refresh_interval": 30,
    "web_port": 8080
}

Changes require service restart: sudo systemctl restart noisy-capture

Microphone → BirdNET-Pi → PulseAudio → Icecast Stream
                                            ↓
                                      Noisy Pi (ffmpeg)
                                            ↓
                              Analyze → Store → Dashboard

Noisy Pi uses ffmpeg to capture audio from BirdNET-Pi's Icecast stream:

• Raw audio capture for FFT spectral analysis
• silencedetect filter for quiet periods
• 256-bin FFT for full 0-24kHz coverage

This approach ensures zero interference with BirdNET-Pi's operation.

Each measurement (every 30 seconds by default) includes:

sudo bash /opt/noisy-pi/uninstall.sh

measurements:
  - id, timestamp, unix_time
  - mean_db, max_db, min_db
  - l10_db, l50_db, l90_db
  - band_0_200, band_200_500, band_500_1k, band_1k_2k
  - band_2k_4k, band_4k_8k, band_8k_24k
  - spectral_centroid, spectral_flatness, dominant_freq
  - silence_pct, dynamic_range
  - anomaly_score, annotation
  - sample_seconds, status
  - spectrogram (BLOB), spectrogram_snapshots, spectrogram_bins

baseline:
  - day_of_week (0-6), hour (0-23)
  - mean_db_avg, mean_db_std, samples

snippets:
  - id, timestamp, measurement_id
  - filename, anomaly_score

• No continuous recording: Only extracted features are stored
• Snippets opt-in: Audio clips are disabled by default
• Local storage: All data stays on your Raspberry Pi
• User control: Delete snippets anytime via dashboard
• Manual annotation: Add context to measurements for review

# Check capture service
sudo systemctl status noisy-capture

# Check Icecast stream is available
ffmpeg -hide_banner -i http://localhost:8000/stream -t 3 -f null - 2>&1 | grep -E "Audio|Duration"

# View capture logs
journalctl -u noisy-capture -n 50

# Check web service
sudo systemctl status noisy-web

# Check PHP
php -v

# Try different port if 8080 is busy
cat /opt/noisy-pi/config/noisy.json | grep web_port

• The baseline needs time to learn (100+ samples)
• Check if actual noise events occurred
• Adjust anomaly_threshold if too sensitive

MIT License - see LICENSE file.

Inspired by BirdNET-Pi.