The recording has one home
Imported media stays in the local workspace. ZoneVoice does not use an implicit cloud fallback for model processing.
Privacy first. Local by design.
Local AI meeting notes.
Your recordings stay yours.
ZoneVoice is free offline transcription software for macOS and Windows. Install the models first, then transcribe recordings and generate meeting notes on your computer. Keep the audio, transcript, and summary together.
audio-recording.m4a 44:51
Transcribe→Separate speakers→Summarize
⌁ Voiceprint vectors encrypted in local SQLite storage
A boundary you can understand
Processing happens where the recording lives.
The desktop app uses local model runtimes for transcription, speaker diarization, and generated note outputs. It does not silently upload media to make the workflow work.
Connecting an AI agent is optional. External clients may send authorized text to their own model providers. See how agent access works.
Website registrations are separate from the desktop product. Submitting an email here never sends a recording, transcript, or voiceprint.
Private AI transcription · Follow the data
Where does each part of the recording go?
Local AI meeting notes depend on the whole workflow. Here is the boundary for the current desktop preview, from an imported file to optional agent access.
| Stage | On your computer | Network or sharing boundary |
|---|---|---|
| Recording / import | An imported copy of your audio or video and a normalized audio file stay in the app’s local storage. The original file stays where you put it. | Import an existing audio or video file. Live system-audio recording is not available in this preview. Audio is not uploaded for processing. |
| Transcription | Whisper processes audio on your computer. Optional speaker separation also runs locally. Timestamped text is saved in the local database. | First-time model and runtime installation needs a download. Missing models do not trigger a cloud transcription fallback. |
| Summary & knowledge | The local Gemma runtime reads transcript text to create notes, tasks, and workspace memory. Saved results and source references stay locally. | No external AI provider is needed for these steps. Generated facts and citations still need your review. |
| Storage & export | Recordings, transcripts, notes, and memory remain in the app’s private data directory. Export creates a file at the location you choose. | Voiceprint vectors are encrypted; this is not a claim that the entire database or every recording is encrypted. A synced export folder may upload files through its own service. |
| Optional MCP | A client on the same computer can read saved text from the workspaces you authorize. Access history stays in ZoneVoice. | The client receives that text and may send it to a cloud model. Revocation stops future reads, but cannot recall earlier copies. |
Recognition needs consent and review
Name a voice after reviewing it. Confident matches can reuse a confirmed name in later recordings; manual corrections take priority.
Useful output should remain verifiable
Summaries, tasks, and mind maps are paired with a transcript and timestamps so the original conversation stays accessible.
About named speakers
Recognize voices. Keep the final say.
For each temporary speaker, ZoneVoice can aggregate clear segments and compare the resulting vector against the local library. A name is automatically applied only when the confidence and separation are strong enough. Otherwise, the app shows a possible match for a person to confirm.
Read how speaker matching works