Out of the box, unottr transcribes, separates speakers, searches and exports without any network call. The optional AI features are the only part that can reach outside your computer, and only if you point them at a hosted provider. If you would rather they did not, a local model server closes that gap.
This post walks through it with Ollama; LM Studio works the same way.
1. Install a model server
Install Ollama and pull a model that fits your machine. A mid-size instruction-tuned model is enough for overviews and Ask; what matters most is that it fits in memory with room to spare for Whisper when both run at once.
ollama pull llama3.1:8bOllama serves an OpenAI-compatible endpoint on localhost by default. LM Studio does the same once you start its local server from the Developer tab.
2. Add the connection in unottr
Open Settings, go to the AI tab and pick Ollama or LM Studio as the provider. Point it at the local server address and choose the model you pulled. unottr shows the connection’s destination and status, so you can see at a glance that it is talking to your own machine.
Consent is stored per connection. Agreeing to use a server on your own laptop is a separate decision from agreeing to use a hosted API, and switching later asks again.
3. Review what would be sent, once
Before the first overview call, unottr shows the exact text it is about to send: transcript segments and speaker names, nothing more. There is a pseudonymize toggle that replaces names with “Speaker A”, “Speaker B” and so on before sending. With a local server the request never leaves the machine, but the review step is the same, so you always know what a connection receives.
Audio, video, frames and project files are never part of a request, to any connection.
4. Generate an overview, then Ask
Open a recording and generate an overview: a title, TL;DR, decisions and tasks, each linked to the moment in the recording it came from. Then try Ask across several meetings. Every answer cites recordings, speakers and timestamps that resolve to real segments; the model cannot invent one.
Local models are slower and less fluent than the largest hosted ones. In practice they are more than good enough for “what did we decide about the launch date”, and the citations let you check the answer in seconds.
What stays where
| Data | Where it goes |
|---|---|
| Recordings, transcripts, speaker voiceprints | Your disk only |
| Search index, Ask threads, notes | Local SQLite database and the folders you choose |
| Overview and Ask requests | The connection you configured; with Ollama or LM Studio, your own machine |
| Whisper and speaker model weights | Downloaded once on first use, then cached |
Automatic overviews are opt-in and off by default. If you later add a hosted connection for a stronger model, that is a separate consent, and the app will show the payload again before the first call.
That is the whole setup. One local service, one connection in Settings, and a meeting library that works the same with the network off.
