🎙️ Home Assistant Voice PE

Your Voice PE, answered by your Companion. Companion Voice connects the Home Assistant Voice Preview Edition  to a Companion that runs on your own hardware, answers with your local model, and remembers what matters to you through Companion Memory.

What it adds to Home Assistant

Companion Voice runs a small Wyoming  server on TCP port 10700. Home Assistant’s built-in Wyoming Protocol integration turns it into three services, which you can use together or on their own:

The simplest setup uses only companion-conversation and keeps Home Assistant’s own faster-whisper and Piper add-ons for speech. No custom components are needed.

How a question travels

The Voice PE hears its wake word on the device. Home Assistant’s Assist pipeline turns speech into text and hands it to the voice bridge, which asks the Companion API for an answer, drawing on your memory first.

flowchart LR
    PE["Voice PE<br/>wake word · mics · speaker"] --> HA["Home Assistant Assist"]
    HA -->|"Wyoming, TCP 10700"| VB["Voice bridge"]
    VB -->|"HTTP 8080"| API["Companion voice API"]
    API --> LLM["Local model<br/>Ollama or Hub inference"]
    API --> SP["Speech server<br/>optional"]
    API <-->|"context · recall · save"| CM["Companion Memory"]

Set it up

cd companion-voice
docker compose -f compose.yml up -d
docker compose -f compose.yml exec ollama ollama pull llama3.2

Connect Home Assistant

Choose how it speaks

Companion Memory

With memory connected, every answer starts by fetching your profile context and the memories most relevant to what you asked, within a 1.5 second budget. They reach the model as reference data, never as instructions. After the answer, the exchange is saved as a conversation from ha-voice-pe, grouped by Home Assistant’s conversation, so it shows up in your memory like any other chat.

• Scope: the key needs only the memory scope. Under Hub, the connect flow creates exactly that. For Docker, create one in Companion Memory under Settings, API keys, as described in Connect Companion Memory.
• Saving is optional: set CI_MEMORY_SAVE_TURNS=false to recall without saving.
• Resilient: if memory is slow or offline, the assistant still answers, without it.

Configuration

The Docker .env from ha.ci.computer sets everything. The variables you are most likely to change:

Check your setup

The last step at ha.ci.computer  asks your server what it is connected to and shows each part: the API, your key, the model, speech and memory. From any computer on your network you can ask the same thing:

curl http://<host>:8080/health
curl -H "Authorization: Bearer $MASTER_API_KEY" http://<host>:8080/v1/status

Under Companion Hub, use the Voice API key from the app’s page in Hub as MASTER_API_KEY, and the address where your Hub exposes the app’s API in place of <host>:8080.

Troubleshooting

Home Assistant cannot connect Wyoming is plain TCP, not a web address. Enter only the IP and port 10700, and make sure no firewall blocks it. From the Home Assistant host, nc -vz <ip> 10700 should succeed.

The Voice PE lights up but never answers Run the check. A failing language model usually means the model is still downloading or was never pulled.

Speech-to-text returns nothing With fully local voice, download the speech models once (the setup page lists the two commands). With the simplest setup, choose faster-whisper and Piper in the assistant instead of the companion services.

Answers do not use my memory The key needs the memory scope, and the address must end in /api when you reach Companion Memory through its gateway. The check shows the exact error.

The web check fails but curl works Browsers block an https page from calling a plain http address on your network. Serve the API over https with Tailscale Serve or a Cloudflare Tunnel, or trust the curl result.

Links

• Guided setup: ha.ci.computer 
• How the Companion devices compare
• Home Assistant Voice PE hardware 
• Wyoming Protocol integration