Where your voice goes

This page follows one spoken sentence from your mouth to the answer, and says what happens to it at each hop. It describes what the shipped software does. Where a hop depends on something you chose, it says so instead of guessing on your behalf.

What Leaves The Device

The wake word never leaves the device. Detection runs locally, on a model that ships inside the app. Nothing is sent anywhere while a device waits for “Hey Jarvis” — that’s the whole point of a wake word.

From the phone app, the hub receives text, not audio. The sound of your voice becomes words on the phone; the words travel to the hub. The hub answers with text, and the phone speaks it. On other clients this depends on the speech plugins installed there — see below.

The link to the hub is encrypted. Client and hub authenticate each other before anything is sent; the client records the hub’s identity on first contact and checks it on every later connection, so a changed hub is an error rather than a silent redirect.

Hop by Hop, on the Phone

1. Waiting. The microphone feeds a local wake-word model bundled in the app. Audio is examined in 80 ms frames and discarded. It isn’t written to storage, and it isn’t sent anywhere.

2. Woken. Once the wake word is confirmed, the app asks Android to turn the next sentence into text.

   The app asks for offline recognition first. If your phone has the offline model for the language you’re using, the sound never leaves the phone.

   If it doesn’t have that model, the app falls back to your phone’s online recogniser — so for that sentence, audio goes to whoever provides speech recognition on your phone, usually its manufacturer or Google. That hop is between you and your phone’s vendor. Thalovant isn’t in it and never receives the audio.

3. Asking. The recognised text goes to your hub over the encrypted link. Audio never does.

4. Answering. The hub replies with text. The phone speaks it using the phone’s own text-to-speech voice. Generating the sound is local.

Note

To keep every sentence on the phone, install your language’s offline speech recognition in Android’s settings. The app prefers it whenever it’s there.

Hop by Hop, on a Speaker or a Computer

Thalovant Voice, the client for rooms and appliances, detects the wake word locally in the same way. What it does next depends on the speech plugins installed on that machine: both recognition and the voice are pluggable, and the default resolves to whatever is installed rather than to a fixed service.

So this hop is yours to know. Check which plugins are configured on that machine before assuming audio stays on it.

Thalovant Voice can also retain recordings for debugging, and doesn’t by default — retention happens only when you set a capture directory.

What the Hub Keeps

Your hub is a piece of software running somewhere you chose: your own machine, or Thalovant’s infrastructure. What it retains (conversation memory, reminders, notes) depends on the skills you install and on the hub’s own settings, not on this page.

Two facts about memory that are worth carrying:

• Memory is per connection. A notebook belongs to the client that wrote it.
• The hub keeps nothing between turns unless a skill stores it. Follow-up questions work because the client sends the session back, not because the hub is recording you.

What This Page Cannot Tell You

Some questions have answers that belong to your workspace rather than to the software, and inventing them here would be worse than leaving them out:

• how long Thalovant-hosted hubs retain logs;
• which sub-processors a hosted hub uses;
• the data-residency options for a hosted hub;
• the published retention period for account and billing data.

Ask support for the current answers, and check status.thalovant.com if you’re investigating something that’s happening right now.