❓ Questions / Help / Support

[bil21071]

Summary

I’m seeing very high false positives when deploying custom wake-word models
(“hilfe” and “adele”) trained with openWakeWord’s 2nd training approach.
Recall is good, but FP rate is high for on-device use like it is >=1 per hour .

---

Environment

• Python: 3.10+
• openWakeWord: 0.6.0
• Hardware:
  • Training: VM
  • Inference: Samsung Galaxy Watch

---

Training Details

• Training approach: 2nd approach from the official training notebook
  (training_models.ipynb)
• Model architecture:
  • Model layer size: 256
• Input shapes:
  • (6, 96)
  • (3, 96)
• Key hyperparameters:
  • recall_weight = 0.2

---

Dataset

Positive samples (original + augmented)

• “hilfe”: ~88,000 samples
• “adele”: ~83,417 samples

Negative samples (~7000 hours total)

• Room impulse responses (RIRs)
• Mozilla Common Voice
• TuDa German speech
• Birds / environmental sounds
• Radio, info content (German & English)

---

Observed Behavior

• Recall: High (≈0.85 and ≈0.77)
• False positives: High
  • Frequent triggers on unrelated speech and background audio
  • Example: random German radio or conversational speech activates “hilfe” / “adele”
  • Custom inference pipeline

---

Expected Behavior

• FP rate similar to pre-trained openWakeWord models
  (e.g. < 1 false trigger per hour on negative-only audio)
• Maintain good recall on positives

---

Question / Help Requested

What is the recommended way to reduce false positives without killing recall?

Any guidance would be greatly appreciated — this is currently blocking on-device deployment.

Thanks!