Voice & AI Integrations

Overview

The home runs a fully local voice pipeline using HA's Assist framework. All components — wake word detection, speech-to-text, text-to-speech, and LLM conversation — run on local hardware with no cloud dependency. HA Cloud is also configured as a fallback/alternative.

Voice Pipeline Architecture

Microphone (HA Voice PE)
       ↓
Wake Word: openWakeWord
       ↓
STT: Whisper (local) or Speech-to-Phrase (local) or HA Cloud
       ↓
Intent / LLM: Home Assistant Assist OR Ollama (qwen3:4b)
       ↓
TTS: Piper (local) or Google Translate or HA Cloud
       ↓
Speaker (HA Voice PE)

---

Hardware

Home Assistant Voice PE

Device: Home Assistant Voice 0919e8 | Manufacturer: Nabu Casa | Model: Home Assistant Voice PE | ESPHome: 2026.3.2 | Area: Bedroom

The dedicated voice satellite device for local Assist.

Entity	Notes
assist_satellite.home_assistant_voice_0919e8_assist_satellite	Satellite control
media_player.home_assistant_voice_0919e8	Audio playback
light.home_assistant_voice_0919e8_led_ring	LED ring indicator
switch.home_assistant_voice_0919e8_mute	Microphone mute
switch.home_assistant_voice_0919e8_wake_sound	Chime on wake
event.home_assistant_voice_0919e8_button_press	Physical button
select.home_assistant_voice_0919e8_assistant_2	Pipeline selector

---

Wake Word Detection

Add-on: openWakeWord v2.1.0 | Wyoming protocol

openWakeWord is an open-source wake word detection library. It processes the audio stream locally to detect activation phrases before sending to STT.

Integration	Entry
wyoming (openWakeWord)	01KSAJTKT9G4SPH8F262Y0R656

Entity: wake_word.openwakeword

---

Speech-to-Text (STT)

Two local STT options + HA Cloud:

Option	Add-on	Version	Notes
Whisper	core_whisper	3.2.0	General-purpose neural STT
Speech-to-Phrase	core_speech-to-phrase	1.4.3	Fast, intent-optimized local STT
HA Cloud	—	—	Nabu Casa cloud fallback

Recommended: Speech-to-Phrase for speed; Whisper for accuracy with free-form speech.

---

Text-to-Speech (TTS)

Option	Add-on	Notes
Piper	core_piper v2.2.2	Local neural TTS (English)
Google Translate	—	Cloud TTS (en-com)
HA Cloud	—	Nabu Casa cloud TTS

Entity: tts.piper, tts.google_translate_en_com, tts.home_assistant_cloud

---

Conversation / LLM

Home Assistant (Assist)

Entity: conversation.home_assistant — built-in intent recognition for HA commands.

Ollama (Local LLM)

Integration: ollama | Host: http://192.168.1.85:11434 | Model: qwen3:4b-instruct | Firmware: 4b-instruct

Ollama runs a local LLM on the home server at 192.168.1.85. The qwen3 model (4B parameters, instruction-tuned) handles natural language conversations and complex commands beyond Assist's built-in intents.

Entity: conversation.ollama_conversation

The Ollama server at 192.168.1.85 must be running for this conversation agent to be available. HA Assist falls back to the built-in agent if Ollama is unreachable.

---

Home Assistant Cloud (Nabu Casa)

Integration: cloud | Status: loaded

Nabu Casa Cloud provides:

• Remote access (no port forwarding required)
• Cloud STT/TTS as fallbacks
• Google Assistant / Amazon Alexa voice control relay (if configured)
• Webhook relay for IFTTT and mobile notifications

---

Notes / Gotchas

• The select.home_assistant_voice_0919e8_assistant_2 entity allows switching the Voice PE between different conversation pipelines (e.g., Assist built-in vs Ollama).
• Whisper accuracy improves with higher model variants — the add-on may be configured with a small model for speed; check the add-on config for model selection.
• Speech-to-Phrase is purpose-built for HA commands and significantly faster than Whisper for home control tasks.
• Google Translate TTS requires internet access; Piper works fully offline.

Last Updated

2026-06-24