Skip to content

Tips for Best Results

Expert tips and best practices for creating high-quality wake word models.

Recording Quality

Environment

Factor Recommendation
Background noise Minimize - record in a quiet room
Echo Reduce - avoid empty rooms with hard walls
Distance Consistent - arm's length from microphone
Device Same device you'll use for detection

Recording Technique

Best Practices

  • Speak at your natural pace and volume
  • Keep recordings 0.5-2 seconds (not too short, not too long)
  • Vary your tone slightly between recordings
  • Include natural variations (morning voice, tired voice)

Common Mistakes

  • Speaking too slowly or deliberately
  • Whispering or shouting
  • Recording in noisy environments
  • Very different recording conditions than deployment

Number of Recordings

Recordings Quality Recommendation
1 Minimum Works with TTS augmentation
3 Good Recommended for most users
5 Better Best accuracy

Wake Word Selection

Ideal Characteristics

Characteristic Good Example Poor Example
Syllables 2-3 ("jarvis", "phoenix") 1 ("hi", "go")
Uniqueness Uncommon words Common words ("computer")
Pronunciation Easy, consistent Hard to say consistently
Phonetic distinction Unique sounds Sounds like many words 
✓ Good choices:
  - jarvis, phoenix, friday, alfred
  - hey buddy, wake up, listen up
  - Custom names: nimbus, orion, echo

✗ Avoid:
  - hi, hello, hey (too common)
  - supercalifragilistic (too long)
  - the, be, okay (appear in speech)

Training Settings

For Best Accuracy

Setting Value
Target Positives 5,000
Negative Ratio 2.0x
Hard Negative Ratio 4.0x
Batch Size 32
Max Epochs 100
Early Stopping 8
All Enhancements Enabled
Setting Value
Target Positives 8,000
Negative Ratio 3.0x
Hard Negative Ratio 6.0x
Batch Size 32
Max Epochs 200
Early Stopping 15
All Enhancements Enabled
Setting Value
Target Positives 3,000
Negative Ratio 1.5x
Hard Negative Ratio 2.0x
Batch Size 64
Max Epochs 50
Early Stopping 5
Enhancements Optional

Hard Negatives

Critical for reducing false positives:

Ratio Trade-off
2.0x Faster training, may have more false positives
4.0x Good balance (recommended)
6.0x Best accuracy, slower training

Reducing False Positives

During Training

  1. Increase hard negative ratio to 5.0x or 6.0x
  2. Enable all model enhancements (attention, SE, TCN)
  3. Use more recordings (5 if possible)
  4. Increase dropout to 0.5-0.6

After Training

  1. Raise the threshold from 0.65 to 0.75-0.85
  2. Test with similar words to identify triggers
  3. Retrain if false positive rate is too high

Common False Positive Causes

Cause Solution
Similar words trigger More hard negatives
Partial utterances trigger Add pure prefix negatives
Background speech triggers More real negatives
Confident on everything Higher dropout, label smoothing

Reducing False Rejections

During Training

  1. Increase target positive samples to 6,000-8,000
  2. Lower dropout to 0.3-0.4
  3. Decrease label smoothing to 0.05
  4. Increase focal alpha to 0.6-0.7

After Training

  1. Lower the threshold from 0.65 to 0.5-0.6
  2. Test in your actual environment
  3. Retrain with more recordings if issues persist

Model Enhancement Options

When to Enable Each

Enhancement Enable When Disable When
Self-Attention Complex wake words, similar sounds Fast training needed
SE Block Fine discrimination needed Memory constrained
TCN Block Local patterns important Simpler wake words
Focal Loss Class imbalance, hard examples Balanced data

Default Configuration

All enhancements are enabled by default for best results. Only disable if:

  • Training is too slow
  • Memory issues occur
  • Simpler model is preferred

Testing Thoroughly

Before Deployment

Test Purpose
Normal speech Baseline detection
Whisper Low volume
Shout High volume
Near/far Distance variation
With noise Background robustness
Similar words False positive check
Different people Speaker variation

Threshold Calibration

  1. Start at the recommended threshold (values in metadata)
  2. Test with 20+ utterances
  3. Count detections and misses
  4. Adjust threshold:
  5. Too many misses? Lower threshold
  6. Too many false triggers? Raise threshold

Troubleshooting

Low Accuracy

Issue Solutions
Accuracy < 90% More recordings, more epochs
F1 < 0.85 Higher hard negative ratio
High train/val gap More dropout, label smoothing

High False Positive Rate

Issue Solutions
FAR > 10% More hard negatives (5.0x+)
Triggers on silence Add noise samples
Triggers on music Add music negatives

High False Rejection Rate

Issue Solutions
FRR > 10% Lower threshold
Doesn't detect user More diverse recordings
Works for TTS but not user Record in actual environment