Micro Wake Word Models for Home Assistant Voice PE

Custom microWakeWord models for on-device wake word detection on Home Assistant voice hardware.

These models run directly on the ESP32-S3 chip — no cloud services, no streaming to a server. Wake word detection happens entirely on the device itself.

Available Models

Wake Word	Language	Model
Nocturna	English	models/nocturna/

Usage with Home Assistant Voice PE

Add the model to your Voice PE's ESPHome configuration by overriding the micro_wake_word section:

substitutions:
  name: home-assistant-voice-XXXXXX
  friendly_name: My Voice PE

packages:
  Nabu Casa.Home Assistant Voice PE:
    github://esphome/home-assistant-voice-pe/home-assistant-voice.yaml

esphome:
  name: ${name}
  name_add_mac_suffix: false
  friendly_name: ${friendly_name}

api:
  encryption:
    key: "YOUR_API_KEY"

wifi:
  ssid: !secret wifi_ssid
  password: !secret wifi_password

micro_wake_word:
  id: mww
  models:
    - model: https://github.com/MorningstarOwl/wake-word-models/raw/main/models/nocturna/nocturna.json
      id: nocturna

Replace the name, friendly_name, and api.encryption.key with the values from your existing Voice PE config.

After flashing, go to Settings → Devices & Services → ESPHome, find your Voice PE, and select Nocturna from the wake word dropdown.

Usage with Other ESPHome Devices

Any ESPHome device with microWakeWord support (ESP32-S3 with PSRAM) can use these models. Add the model to your micro_wake_word configuration:

micro_wake_word:
  models:
    - model: https://github.com/MorningstarOwl/wake-word-models/raw/main/models/nocturna/nocturna.json

See the ESPHome microWakeWord documentation for full configuration options.

Tuning

If the wake word triggers too easily or not reliably enough, you can override the detection parameters in your ESPHome YAML without re-training:

micro_wake_word:
  id: mww
  models:
    - model: https://github.com/MorningstarOwl/wake-word-models/raw/main/models/nocturna/nocturna.json
      id: nocturna
      probability_cutoff: 80%       # Lower = more sensitive (default from manifest: 97%)
      sliding_window_size: 8        # Higher = smoother detection, slightly more latency

How These Models Were Trained

Models are trained locally using the microWakeWord Trainer Docker container, which automates the full pipeline:

1. Synthetic speech samples are generated using Piper TTS
2. Samples are augmented with room simulation, background noise, and speed variation
3. A streaming MixConv neural network is trained with TensorFlow
4. The model is quantized and exported to TFLite for deployment on microcontrollers

Personal voice recordings can optionally be added to improve detection accuracy for specific speakers.

Related Resources

• microWakeWord — The training framework
• ESPHome micro_wake_word docs — Configuration reference
• Home Assistant Voice PE — Official Voice PE firmware
• ESPHome wake word models — Official model repository

License

This project is licensed under the Apache License 2.0.