Hotword

Custom wake word detection. Like "Hey Google" or "Hey Siri" — but yours.

Record your wake phrase, train in under 2 minutes, deploy. No cloud, no subscription, runs on-device.

Your wake word, visualized in 3D

Better than Porcupine. Better than OpenWakeWord.

F1=1.000. Zero false positives, 100% recall. Trains in 90 seconds on a laptop. No license fees, no cloud, no per-device limits.

How it works

1. Record your wake phrase (~1s clips)
2. Embed using a speech embedding model (16 temporal frames x 96 dims)
3. Train a lightweight ConvAttn classifier (~130K params) with attention pooling
4. Detect in real-time with edge-based triggering (no duplicate fires)

The classifier learns temporal patterns in the embedding space — it knows when each part of your phrase should appear, not just that speech is present.

Quick start

# Record positive samples (say your wake phrase)
make record-pos

# Record negative samples (say other things)
make record-neg

# Embed and train
make embed
make train

# Evaluate
make eval

# Run live detection
make detect

Architecture

audio → mel spectrogram → speech embedding (ONNX) → [16 × 96]
  → Conv1D (temporal features) → attention pooling → sigmoid → wake/not-wake

Edge-based detection: fires once per utterance, re-arms when confidence drops. No duplicate triggers, no cooldown hacks.