Audio Event Detection

Audio Event Detection

Ava Pro includes an on-device audio event detection system that listens for common household sounds and reports them to Home Assistant. The entire pipeline runs locally — no audio leaves the device, no cloud services involved.

Audio event detection operates on a dedicated processing thread, completely isolated from wake word detection and voiceprint processing. This means event detection never adds latency to wake detection, and vice versa.

Compatible with Android 5-16.

---

Performance

Metric	Value
Overall accuracy	98.2%
False positive rate	1.3%
Detection latency	< 200ms (single window)
CPU impact (idle room)	Near zero — inference skipped on silent windows
CPU impact (active detection)	Single-thread TFLite, adaptive interval 1-2s

---

How It Works

The system captures 1-second audio windows from the microphone and runs a lightweight TFLite classifier. Each window is analyzed independently — no continuous recording, no streaming.

The processing pipeline:

1. Window capture — 1 second of 16kHz mono audio
2. Quality gate — silent or dead audio is discarded before inference
3. Feature extraction — 40-band mel-frequency spectrogram (99 frames × 40 filters)
4. Classification — 7-class TFLite model outputs confidence scores
5. Confidence gate — only windows above the sensitivity threshold pass
6. Speech confusion guard — if speech is a strong runner-up, non-speech events are suppressed
7. Confirmation — borderline detections require a second confirming window within 4-6 seconds
8. Publish — confirmed event label is sent to the Home Assistant sensor

The classify interval adapts to device performance: fast devices run inference every 1 second, slower devices every 1.5-2 seconds. Low-RAM devices also receive slightly stricter quality gates to reduce false positives.

---

Detection Types

Seven sound categories are supported:

Type	Description	Example Use Case
Alarm	Smoke and CO alarm tones	Fire safety monitoring
Noise	Crying, distress, or other everyday household sounds (renamed from "Baby crying" in 0.6.7)	Baby monitor, general alert
Cough	Coughing sounds	Health tracking
Doorbell	Doorbell and chime sounds	Visitor notification
Glass breaking	Shattering or impact on glass	Security automation
Siren	Emergency vehicle and outdoor sirens	Emergency alert
Speech	Talking and conversation	Presence detection

By default, only speech monitoring is enabled. Additional types can be toggled in Settings. At least one type must remain enabled — the last type cannot be turned off.

---

Sensitivity

Three sensitivity presets control how easily events trigger:

Preset	Description	Best For
Fewer false alerts (Conservative)	Stricter confidence and quality gates	Quiet homes, low false positive tolerance
Balanced (Default)	Recommended threshold for most rooms	General household use
Catch more events (Sensitive)	Looser gates, may increase false positives	Critical monitoring, noisy environments

Sensitivity affects four internal thresholds:

• Minimum confidence score to accept a detection
• Speech runner-up suppression level
• Feature quality floor
• Single-window strong detection threshold

---

Home Assistant Entity

sensor.your_device_name_audio_event

The sensor reports the detected event label (e.g., alarm, baby_cry, doorbell), or idle when no event is active.

After an event is published, the sensor holds the label for a configurable duration (5-60 seconds, default 30 seconds) before returning to idle. The display duration can be adjusted in Settings and takes effect immediately.

0.6.7: The former "Baby cry" label is now the more everyday "Noise," covering crying or distress sounds and clearly warning that this category may produce more false alerts. This avoids presenting it as a dedicated baby monitor.

Automation Example

# Notify when a doorbell is detected
trigger:
  - platform: state
    entity_id: sensor.your_device_name_audio_event
    to: "doorbell"
action:
  - service: notify.mobile_app
    data:
      message: "Doorbell detected at home"

# Turn on lights when baby crying is detected
trigger:
  - platform: state
    entity_id: sensor.your_device_name_audio_event
    to: "baby_cry"
action:
  - service: light.turn_on
    target:
      entity_id: light.nursery

---

Setup

1. Go to Settings → Voice Config → Audio events
2. Turn on Audio event detection
3. Set Alert status display duration (how long the sensor holds the event label)
4. Under Listen for, toggle the sound types to monitor
5. Under Sensitivity, choose the preset that fits the environment
6. The audio_event sensor appears in Home Assistant

---

Use Cases

Scenario	Recommended Types	Sensitivity
Baby monitor — notify parents when baby cries	Noise	Catch more events
Fire safety — alert on smoke alarm tones	Alarm	Catch more events
Security — detect glass breakage	Glass breaking	Fewer false alerts
Doorbell notification while away	Doorbell	Balanced
Presence detection — know when someone is talking	Speech	Balanced
Emergency siren monitoring	Siren	Catch more events
Cough tracking for health logs	Cough	Balanced

---

Note: Audio event detection is a convenience feature, not a certified safety or medical monitoring system. Some low-end devices may produce false positives. Critical safety applications should always use dedicated certified detectors.

---

Back to Voice Control