Skip to main content
Sound Security Guard monitors audio for a single active scenario at a time using the pretrained Audio Spectrogram Transformer (MIT/ast-finetuned-audioset-10-10-0.4593). The model outputs 527 AudioSet class probabilities (multi-label sigmoid). No fine-tuning is required—scenarios map to existing AudioSet label strings.

​
Strict scenario isolation

Only labels belonging to the selected scenario are evaluated. For example, with Glass Break active, a passing police siren is ignored even if the model detects it strongly.

​
Scenarios (sub-modalities)

ScenarioMonitored AudioSet labels (substring match)
Glass Break / Falling ObjectsGlass, Shatter, Smash crash, Thump, Crash, Bang
Constant Alarm (Siren)Alarm, Car alarm, Smoke detector, Siren, Ambulance (siren), Police car (siren)
Help! / ScreamingScreaming, Yell, Shout, Children shouting
CustomUser-defined comma-separated label substrings

​
Parameters

ParameterDefaultDescription
confidence_threshold0.40Alert when any scenario label exceeds this probability
output_buffer_presetspeech-to-textAnalysis window size (4 s recommended)
output_buffer_size_s4.0Custom window length (seconds, min 1 s)
event_cooldown_s5.0Minimum seconds between alerts

​
Input / output

Input: Same as VA—audio adapted to PCM S16LE int16 @ 16 kHz mono. Interim output (window filling / no alert): 
{
  "event_detected": false,
  "event_confidence": 0.12,
  "event_label": "",
  "event_type": "glass_break",
  "active_scenario": "glass_break",
  "sample_rate_hz": 16000,
  "channels": 1
}
Alert output: 
{
  "audio": "<numpy int16 — analysis window>",
  "event_detected": true,
  "event_confidence": 0.67,
  "event_label": "Glass",
  "event_type": "glass_break",
  "active_scenario": "glass_break",
  "start_timestamp_sec": 8.0,
  "end_timestamp_sec": 12.0,
  "sample_rate_hz": 16000,
  "channels": 1
}
Wire event_detected to a send_alert node or conditional gate (see Alerts).

​
Analysis window

Audio accumulates until the analysis window is full (default 4 s), then AST runs inference. The buffer advances by 50% hop (half-window overlap) so events near window boundaries are not missed. After an alert, a cooldown suppresses duplicate notifications.

​
Example workflow

Audio Track → Audio Assistant (SSG, glass_break) → Send Alert

​
Edge install

pip install "cyberwave[ml-aed]"
First inference downloads the AST weights from Hugging Face.