In-Car SOS with On-Device Vector Search

Detecting distress sounds passively using Qdrant Edge — no cloud, no raw audio upload, no manual SOS button.

---

Overview

This system continuously monitors in-car audio and detects distress sounds (screams, glass breaking, tire screeches, collisions) using on-device vector similarity search powered by Qdrant Edge. When distress is confirmed, an alert is sent via Telegram.

USB Mic → YAMNet (1024-d embeddings) → Qdrant Edge ANN Search → Telegram Alert

Privacy guarantee: Raw audio never leaves the device. Only outbound call is the Telegram alert.

---

Architecture

---

Quick Start

1. Install dependencies

python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

2. Configure environment

cp .env.example .env
# Edit .env with your Telegram bot token and chat ID

3. Index sound library (one-time)

python scripts/indexing_sounds.py

This downloads the ESC-50 dataset, incorporates any custom samples from data/custom/, generates YAMNet embeddings for all distress + negative sounds (with engine noise augmentation for custom samples), and builds the local Qdrant Edge shard (~6 MB).

4. Test the pipeline

python scripts/test_detection.py

Verifies Telegram connectivity, loads the shard, and runs synthetic audio through the detector.

5. Start the live detector

python main.py

---

Configuration

All tunable parameters live in .env:

Variable	Default	Description
TELEGRAM_BOT_TOKEN	—	Your bot token from @BotFather
TELEGRAM_CHAT_ID	—	Your personal or group chat ID
SIMILARITY_THRESHOLD	0.80	Cosine similarity cutoff (0–1)
SMOOTHING_HITS_REQUIRED	3	Hits needed within window to alert
SMOOTHING_WINDOW_SECONDS	5	Time window for hit counting
ALERT_COOLDOWN_SECONDS	30	Min seconds between consecutive alerts
CHUNK_DURATION_SECONDS	1	Audio window size
OVERLAP_SECONDS	0.5	Overlap between windows

Advanced False-Positive Filtering (src/detector.py)

To prevent false alarms in noisy environments, the system uses custom per-class overrides:

• Amplitude Gating: The system strictly ignores low-volume ambient noises (like AC or engine hums) before they even reach the AI model.
• Strict Mode (car_horn, siren): Broadband noises like wind can mimic sirens/horns. These require a very high similarity score (e.g., 0.90 or 0.96) to trigger.
• Sensitive Mode (scream, gunshot): Since emergency human screams and gunshots can be brief or muffled, they are highly sensitive (score 0.80 to 0.85) and require fewer consecutive hits to alert you instantly.

---

Sound Classes

Alert sounds (trigger SOS):

• scream / crying → severity: high / medium (includes custom datasets)
• glass_break → severity: high
• collision / gunshot → severity: high (includes custom datasets)
• siren → severity: high
• car_horn → severity: medium

Negative sounds (do NOT trigger):

• Normal speech, music, engine noise, ambient sounds

---

Tech Stack

Component	Technology
Vector database	Qdrant Edge (qdrant-edge-py)
Audio embedding	YAMNet via TF Hub (1024-d)
Audio preprocessing	librosa
Audio capture	sounddevice
Alert delivery	Telegram Bot API
Datasets	ESC-50, custom .wav/.mp3 folders

---

Screenshots

Project Structure

qdr-edge/
├── main.py                    # Entry point
├── requirements.txt
├── .env.example               # Config template
├── scripts/
│   ├── index_sounds.py        # One-time indexing pipeline
│   └── test_detection.py      # End-to-end test
└── src/
    ├── config.py              # Environment config
    ├── audio_capture.py       # Real-time mic capture
    ├── preprocessor.py        # Audio normalization
    ├── embedder.py            # YAMNet wrapper
    ├── vector_store.py        # Qdrant Edge interface
    ├── detector.py            # Detection + smoothing
    └── alerter.py             # Telegram + console alerts

---

References

• Qdrant Edge Docs
• YAMNet on TF Hub
• ESC-50 Dataset
• Telegram Bot API