How I connected Cosmo to Home Assistant Voice speakers so my AI assistant can talk to me through the house
Beyond the Screen
In my last post, I introduced Cosmo — my 24/7 AI assistant running on OpenClaw in my homelab. Cosmo manages my notes, monitors services, tracks bookmarks, and handles all sorts of background automation. But every interaction happened through text — Telegram messages, terminal commands, web chat.
I wanted something more ambient. Something where Cosmo could just tell me things — a morning briefing while I’m getting ready, an alert when something goes wrong in the lab, or just a friendly notification that a long-running task finished.
The answer was already sitting on my nightstand: Home Assistant Voice.
The Hardware
I have two Home Assistant Voice satellites in the house — ESPHome-based smart speakers with built-in wake word detection:
| Speaker | Location | Wake Word |
|---|---|---|
| HA Voice PE #1 | Master Bedroom | ”Hey Jarvis” |
| HA Voice PE #2 | Gaming Room | ”Hey Jarvis” |
These are proper bidirectional voice devices. They can listen (via Assist pipelines) and they can speak (via TTS services and the announce feature). That second capability is what I was after.
The Integration
Home Assistant exposes everything as REST API calls. Since Cosmo already had access to my HA instance over Tailscale, the integration was surprisingly straightforward.
TTS: Making It Talk
Home Assistant supports multiple TTS engines. I tested three:
1. Nabu Casa Cloud TTS (recommended — best quality)
curl -s -X POST \
-H "Authorization: Bearer $HASS_TOKEN" \
-H "Content-Type: application/json" \
"$HASS_URL/api/services/tts/speak" \
-d '{
"entity_id": "tts.home_assistant_cloud",
"media_player_entity_id": "media_player.bedroom_speaker_media_player",
"message": "Good morning Brian. You have 3 new Raindrop saves and no alerts."
}'2. Google Translate TTS (free, robotic)
curl -s -X POST \
-H "Authorization: Bearer $HASS_TOKEN" \
-H "Content-Type: application/json" \
"$HASS_URL/api/services/tts/speak" \
-d '{
"entity_id": "tts.google_translate_en_com",
"media_player_entity_id": "media_player.bedroom_speaker_media_player",
"message": "Hello from the free tier."
}'3. Legacy Cloud Say (still works, less flexible)
curl -s -X POST \
-H "Authorization: Bearer $HASS_TOKEN" \
-H "Content-Type: application/json" \
"$HASS_URL/api/services/tts/cloud_say" \
-d '{
"entity_id": "media_player.bedroom_speaker_media_player",
"message": "This is the legacy method."
}'The Nabu Casa Cloud TTS is the clear winner — natural-sounding voices with good cadence. Worth the Home Assistant Cloud subscription on its own.
Volume Control
Nobody wants to get blasted by a notification at 2 AM:
curl -s -X POST \
-H "Authorization: Bearer $HASS_TOKEN" \
-H "Content-Type: application/json" \
"$HASS_URL/api/services/media_player/volume_set" \
-d '{
"entity_id": "media_player.bedroom_speaker_media_player",
"volume_level": 0.25
}'How Cosmo Uses It
With the REST API mapped out, I gave Cosmo the ability to speak through any speaker in the house. Here’s how it fits into the workflow:
Proactive Notifications
Cosmo runs hourly heartbeat checks — monitoring services, checking for new bookmarks, verifying backups. If something needs my attention, instead of just logging it or sending a Telegram message, it can announce through the nearest speaker:
“Hey Brian, Tailscale went offline on the Proxmox node. You might want to check that.”
Task Completion Alerts
When I kick off a long-running background task (a subagent doing research, a large file sync, a deployment), Cosmo can announce when it’s done:
“Your blog deployment finished successfully.”
Morning Briefing
A scheduled summary of what happened overnight — new saves, completed tasks, any alerts that fired. Delivered through the bedroom speaker while I’m getting ready.
The Architecture
Here’s how all the pieces connect:
┌──────────────┐ ┌──────────────┐ ┌─────────────────┐
│ OpenClaw │────▶│ Home │────▶│ HA Voice │
│ (Cosmo) │ API │ Assistant │ TTS │ Speakers │
│ │ │ │ │ 🔊 Bedroom │
│ Lab Server │ │ (Tailscale) │ │ 🔊 Gaming Room │
│ │ │ │ │ │
└──────────────┘ └──────────────┘ └─────────────────┘Everything communicates over my Tailscale network — encrypted, no ports exposed to the internet. The HA Voice speakers connect to Home Assistant locally over WiFi, and Cosmo reaches HA via its Tailscale address.
Tips If You’re Building This
Start with tts/speak — it’s the modern service call and gives you the most control over which TTS engine to use.
Set a sane default volume. I keep the bedroom speaker at 25% and only bump it for important alerts. Your future self at 3 AM will thank you.
Use the announce feature for time-sensitive messages. HA Voice speakers support interrupting whatever’s playing to deliver an announcement, then resuming.
Rename your devices. Out of the box, HA Voice speakers have generic serial-based names. Rename them in Home Assistant to something human-readable like bedroom_speaker — it makes automations much easier to reason about.
Test all three TTS methods. Google Translate is free but sounds robotic. Nabu Casa Cloud is natural but requires a subscription. Find what works for your use case.
What’s Next
The obvious next step is making this bidirectional — using the “Hey Jarvis” wake word to trigger Cosmo directly through the speakers, not just HA’s built-in Assist pipeline. Imagine walking into a room and asking your AI assistant a question without reaching for your phone.
I’m also exploring time-aware volume scheduling (quieter at night, louder during the day) and room-aware notifications (speak through whichever speaker is closest based on motion sensors).
The home is becoming an interface. The screen is becoming optional.