Weighted Top-K Audio Detector Training (Late-fusion)
Settings
Booting…
idle
Training
Source
Sender (WebRTC)
Microphone (Receiver)
File
Folder
Choose file…
Choose folder…
Start
Model
Model
Recompute Weights
Clear Model
Export Model
Pick an existing model or choose “New…” to create one. Train (mic/file/folder/sender), then click “Recompute Weights”.
Quiet calibration
Calibrate thresholds
Start
Duration
20s
Thresholds
OFF
0.25
ON
0.35
Keep the room quiet. (Uses the selected source: Mic or Sender.)
Live audio detection
Detection
Start
Uses the active model & thresholds and the selected source (Mic or Sender).
Live top classes
Uses the
Top-K
setting. Excluded classes are grey; we extend to show Top-K non-excluded.
Weights
Live dynamic weights (unsaved preview)
Top-K normalized weights recomputed on every training inference. Not stored.
Active model weights (Top-K, normalized)
What detection uses (until you click “Recompute Weights”).
Fused score over time
Window (s)
fused (raw)
fused (EMA)
thresh ON
thresh OFF
Logs
Display logs
Log inference
Log ALL classes
Top-N
Autoscroll
Clear Logs
Export JSON
Settings
Help
×
Sample Rate (target Hz)
Frame length (s)
Hop (s)
Per-class EMA τ (s)
Fused score EMA τ (s)
Min ON hold (ms)
Quiet calibration (s)
OFF delta (below ON)
Top-K classes
Hybrid prior (kPrior)
Hybrid half-life (s, training EMA)
Max results from MP
Ignore classes (training)
Penalty map (inference)
Reset all defaults