Testing Models¶

How to test your trained wake word models with real-time audio.

Overview¶

WakeBuilder provides real-time testing through your microphone. Speak your wake word and see instant detection feedback.

Real-time Testing¶

Accessing the Test Page¶

Navigate to Test in the navigation bar
Or click "Test Model" on a model card
Or click "Test" after training completes

Testing Interface¶

The real-time testing interface shows:

Model selector: Choose which model to test
Detection indicator: Visual feedback when wake word detected
Confidence meter: Current detection confidence level
Threshold slider: Adjust detection sensitivity
Audio visualizer: Real-time waveform display

How to Test¶

Select model: Choose from the dropdown
Grant microphone access: Click "Allow" when prompted
Start listening: Click the "Start" button
Speak: Say your wake word naturally
Observe: Watch for the detection indicator

Detection Feedback¶

Indicator	Meaning
Green pulse	Wake word detected
Confidence bar fills	Approaching threshold
No response	Not detected

Adjusting Sensitivity¶

Use the threshold slider to fine-tune detection:

Threshold	Behavior
0.3	Very sensitive - more triggers, possible false positives
0.5	Balanced - good for most use cases
0.65	Default - recommended threshold
0.8	Conservative - fewer false positives, may miss some

Finding Your Threshold

Start at the default (0.65) and adjust based on your experience:

Too many false triggers? Increase threshold
Missing detections? Decrease threshold

WebSocket Protocol¶

API Endpoint¶

Real-time testing uses WebSocket:

ws://localhost:8000/api/test/ws/{model_id}

Protocol¶

Client → ServerServer → Client

{
  "type": "audio",
  "data": "<base64 encoded audio>",
  "sample_rate": 16000
}

{
  "type": "detection",
  "detected": true,
  "confidence": 0.87,
  "threshold": 0.65,
  "model_id": "jarvis",
  "timestamp": "2026-01-06T15:30:00Z"
}

Audio Format Requirements¶

Parameter	Value
Sample rate	16000 Hz
Channels	Mono (1)
Format	PCM 16-bit or Float32
Chunk size	~1 second

Understanding Results¶

Confidence Score¶

The confidence score (0-1) indicates how certain the model is:

Score	Interpretation
0.9+	Very confident detection
0.7-0.9	Likely match
0.5-0.7	Uncertain
0.3-0.5	Probably not the wake word
<0.3	Definitely not the wake word

Why Confidence Varies¶

Normal factors affecting confidence:

Speaking style: Casual vs. clear pronunciation
Distance: Close vs. far from microphone
Background noise: Quiet vs. noisy environment
Speaker: Your voice vs. synthesized voices

Testing Best Practices¶

For Accurate Testing¶

Do This

Test in your target environment (where you'll actually use it)
Test with natural speech, not exaggerated pronunciation
Test with background noise if your environment has it
Test with different distances from the microphone

Avoid This

Testing only in perfect silence
Over-enunciating the wake word
Testing at the same volume every time
Ignoring edge cases

Test Scenarios¶

Scenario	Purpose
Normal speech	Baseline performance
Soft speech	Lower volume detection
With music	Background noise robustness
From distance	Microphone sensitivity
Similar words	False positive rate
Rapid repetition	Continuous detection

Troubleshooting¶

Model not detecting my wake word

Lower the threshold to 0.4-0.5
Speak closer to the microphone
Check audio levels aren't too low
If still failing, retrain with more recordings

Too many false positives

Raise the threshold to 0.75-0.85
Retrain with higher hard negative ratio
Avoid common words as wake words

Detection is laggy

Check CPU/GPU usage - may be under load
Close other applications using the microphone
On GPU, ensure CUDA is enabled

Microphone not working

Check browser permissions for microphone access
Ensure microphone is not muted
Try a different browser (Chrome recommended)
Check system audio settings