Owner-only daemon socket
The Unix socket is created mode 0600 inside a mode-0700 runtime directory, restricting access to the current user.
Precise boundaries, not absolute slogans
Local speech. Your chosen model endpoint. No AverVOX telemetry.
AverVOX does not receive your microphone audio, transcriptions, conversations, or usage data. STT and TTS run locally; in Converse, text goes only to the LLM endpoint you configure.
Local STT/TTS/VADNo analytics SDKOffline license verificationOwner-only bridge socket
What crosses a boundary
The endpoint choice determines the conversation-data path.
A local model can keep the entire speech-and-LLM loop on one machine or LAN. A remote model provider necessarily receives the text you send to it under that provider’s terms.
| Data or operation | Processed by AverVOX locally | Sent to AverVOX | May reach your LLM endpoint |
|---|---|---|---|
| Microphone audio | ✓ | Never | Never by AverVOX |
| Speech transcription | ✓ | Never | Yes, only in Converse |
| Selected text for TTS | ✓ | Never | No |
| LLM response text | ✓ for local TTS | Never | Originates from that endpoint |
| Product usage analytics / crash telemetry | Not collected | Never | No |
| License verification | Offline signature check | No license-server call | No |
Local storage
Plain-text history is a feature you can see and control.
Pro can save conversation transcripts and daily dictation logs locally. The configuration documentation identifies their locations and allows the features to be disabled.
- Config
~/.config/avervox/config.yaml- Transcripts
~/.local/share/avervox/transcripts/- Dictation logs
~/.local/share/avervox/dictation/- Bridge socket
$XDG_RUNTIME_DIR/avervox/bridge.sock
Bridge security
Local integrations use narrow, inspectable interfaces.
The developer bridge is designed to avoid exposing a speech service on the network by default.
Protected outputs
Synthesis output files are created mode 0600. Integrations use owner-only temporary locations for text and audio handoff.
Safer process arguments
For sensitive text, the documentation recommends stdin or a file because command-line arguments may be visible in a process listing.
Website privacy
No tracking-cookie funnel required.
The public site can measure success through aggregate server and checkout records without placing third-party behavior trackers in the visitor’s browser.
- No product or website analytics cookies
- Normal server logs may be retained for operational security
- Stripe processes purchase and payment details
- YouTube is loaded only after a visitor opens the demo
- The demo embed uses the privacy-enhanced youtube-nocookie domain
Privacy questions
Make the configuration choice explicit.
Can I honestly say nothing ever leaves the machine?
Only when the configured LLM endpoint is also local. The accurate general claim is that speech processing stays local and conversation text goes only to the endpoint the user chooses.
Does AverVOX call a license server?
Current Pro documentation states that license activation is verified offline using a cryptographic signature, with no recurring license-server call.
Are transcripts always saved?
Pro supports automatic conversation transcripts and daily dictation logs, stored locally. Current documentation says both can be disabled in configuration.
Does the warm daemon open a TCP port?
No. It uses a local Unix-domain socket. The separate Pro LAN server is an explicit network feature and should be configured only on a trusted network.
Where is the binding policy?
The formal privacy policy remains the governing public statement. This page is the product-oriented explanation of the same data boundaries.
Start with the workflow
Keep speech local—and keep the endpoint decision visible.
Start with OSS, inspect the source and network behavior, and choose a local model when the entire conversation must remain on your machine or LAN.