nagisanzeninz/greet-wakeword-v0

Greet Wake Word

Custom INT8 TFLite Micro wake-word detector for ESP32-S3. Fires on:

• hey
• hi
• hello

Metrics

Metric	Value
FRR (positives)	0.00%
FAR (bulk negatives, /hour)	0.00
Detection threshold	0.850
Model size	60.0 kB
Mel features	40 bins, 25 ms hop, 194 frames

Per-bucket FAR:

Bucket	FAR
(no buckets)	—

Use with ESPHome

micro_wake_word:
  microphone: tofu_mic
  vad:
  models:
    - model: greet
      probability_cutoff: 0.850
      sliding_window_size: 5
      url: https://huggingface.co/nagisanzeninz/greet-wakeword-v0/resolve/main/greet-wakeword-v0.tflite

Full ESPHome example in esphome.yaml.

Training recipe

Trained with the customWakeWord pipeline:

1. ~10000 synthetic positives across 3 trigger phrases (Piper + Kokoro + MeloTTS + Parler-TTS, 200+ distinct voices).
2. ~~2500 hand- and LLM-curated hard-negatives covering 5 collision categories.
3. ~300 hours of bulk negative audio sampled from MUSAN, DEMAND, Common Voice, AudioSet.
4. audiomentations chain: RIR convolution + additive noise + codec degradation + pitch/speed jitter.
5. microWakeWord streaming-Inception architecture, INT8 post-training quantized for tflite-micro on ESP32-S3.

See the repo README for replicability.

License

Apache 2.0. Built from open-source, permissively-licensed components only — safe to redistribute commercially.

Citation

@misc{greet-wakeword-v0,
  title  = { greet Wake Word v0: custom keyword spotter for ESP32-S3 },
  year   = { 2026 },
  url    = { https://huggingface.co/nagisanzeninz/greet-wakeword-v0 },
  note   = { Apache 2.0. Built with the customWakeWord toolkit. }
}