From a folder of recordings to answers you can check.

01

Drop a recording in a folder

Any new file in a watched folder is picked up, probed, transcribed and diarized on its own. There is no import step and no start button. The library and a persistent queue shelf show what is finished, what is running and how long it has left.

Every stage is a persisted queue state: discover, probe, extract to 16 kHz mono audio, detect speech, transcribe in roughly 30 second chunks, diarize, merge and index. A crash resumes at a boundary instead of starting over, and compute runs in a separate process so a native crash kills one job, not the app.

02

Speakers that stay themselves

On Linux, automatic speaker separation uses Sortformer on a Vulkan GPU when the runtime is present, and finds up to four speakers without asking for a count. A CPU path handles fallback, separate microphone tracks and exact counts from 1 to 20. Apple Silicon uses the CPU path for this stage.

Name a speaker once and later recordings recognise them by voiceprint. When the split is wrong, merge two speakers, reassign one segment, or run speaker identification again without transcribing the file again.

03

Search everything, land on the second it happened

Full-text search covers every transcript. A hit opens the recording at that timestamp, and a hit inside an AI overview opens the overview tab.

Select a transcript turn to correct its text; search, copy, exports and future AI requests use the correction while timestamps and speakers stay intact. Terminology rules fix recurring names and acronyms in new and existing transcripts, and they are reversible.

04

An AI overview, when you want one

Optional and per recording: a title, a TL;DR, topic sections, decisions and a task list split into yours and everyone else's. Every line links back to the second it came from.

The model cannot invent a timestamp. It cites segment ids, ids resolve to milliseconds in one place, and a citation that does not resolve is dropped. Regenerating keeps tasks you edited, checked or dismissed.

05

Ask across meetings

Ask searches transcript segments, fresh overviews and current tasks, then calls the model you selected. Scope a thread to recordings, people or meeting dates; the scope is pinned when the thread starts so a follow-up cannot silently search different meetings.

Each answer block must cite retrieved sources. Unknown citations are dropped, and a valid one shows the recording, speaker, date, timestamp and excerpt, and opens the video at that moment. Threads stay in the local database.

06

Privacy by construction

There is no network call in the transcription or diarization path. Recordings are watched, transcribed, diarized, searched and exported on your machine. The app fetches model weights on first use, then caches them.

AI overviews and Ask are the exceptions, and nothing happens until you add a connection and start a request. Consent is stored per connection, and before the first overview call unottr shows the exact text it is about to send. Audio, video, frames and project files are never sent. Pointing the app at Ollama or LM Studio keeps the request on your machine.

Bring your own model: Ollama, LM Studio, OpenAI, Anthropic, Mistral, OpenRouter, another OpenAI-compatible endpoint, or an installed Claude Code, Codex or Antigravity CLI.

07

Exports and notes

Export or copy any transcript as txt, json, srt or vtt. Save Markdown notes to an existing folder and edits to transcripts, titles, bookmarks, tasks and speakers update the mirror. If you edit a note outside unottr, automatic updates preserve it.

Personal notes sit beside a recording and survive summary regeneration. Library backups are verified, and restoring one asks for confirmation first.

08

Recording on Linux

Choose Record, check the preview and the microphone and computer-audio meters, then start. Video is compressed into MP4 while recording, at native resolution and 15 fps by default, with resolution and frame-rate controls before you start. Pause or stop from the app, tray or a configured shortcut.

Transcription starts after saving, and interrupted recordings can be recovered from the Record dialog.