1. Separate the stages
A voice workflow can include recording, speech recognition, punctuation, filler removal, rewriting, history, and synchronization. A product can run one stage locally and another in the cloud.
Ask where each stage runs instead of accepting one umbrella word such as “AI” or “private.”
- Where is raw microphone audio processed?
- Where is the first transcript produced?
- Where are cleanup and rewriting performed?
- Is any transcript history synchronized or retained?
2. Treat retention and processing as different questions
A zero-retention cloud mode still sends audio to a server for processing; it changes what happens after processing. A local app can still make narrow network connections for model downloads, licensing, updates, or optional services.
A credible product should disclose both the sensitive data path and the supporting network connections.
3. Match the architecture to the material
Cloud services can provide cross-device access, broad language support, team features, and continuously updated models. Local software can reduce the data path and keep working after setup when the network is unavailable.
The right choice can differ between a public brainstorming note and a confidential legal, medical, financial, or source-code discussion.
4. Read the exception list
Look for the phrases that qualify the headline: first-use model download, account authentication, privacy mode, sync setting, model-improvement opt-in, license verification, or unsupported language.
IraVoice discloses its model download and Gumroad license verification. It also discloses a separate, explicit choice: when Codex or Claude Code is selected as the Prompt destination, the finished specification is submitted to that agent. Microphone audio is not part of any of those connections.