The meeting, on record.

Records any call on your Mac. Transcribes on your machine. Turns the talk into requirements, documents and tickets.

Free. macOS 14.2 or later on Apple silicon, Windows 10 or later. Source on GitHub.

The meeting page in Dictum: the transcript on the left, the recap and the requirements on the right

Record

Anything that makes a sound.

Dictum records what comes out of your speakers and what goes into your microphone, on two separate tracks. Teams, Zoom, Meet, a browser tab, a phone on the desk. No host permission. No plan tier. No bot joining the call.

Two tracksWho said what stays clear even when people talk over each other.

Pause and resumeStep out of the call; the recording waits.

Stops itselfTen minutes of silence after the call ends and it stops and transcribes.

Never loses audioA crash or a force quit keeps every second recorded so far.

Private

Transcribed on your Mac. Nothing leaves.

Speech recognition runs on your machine with the built-in speech model, the best there is for Russian, plus a multilingual one. Recaps, requirements and answers come from a local model, Gemma 4 or Qwen 3, that Dictum downloads and runs for you. Mark a project local-only and the cloud is simply not an option for it.

Works offlineOn a plane, at a client site, behind a corporate firewall.

Your key, if you want itClaude for the heavy lifting, billed to you, never on by default.

One folderRecordings, transcripts, models. Move it, back it up, delete it.

Screens

Every screen, both looks.

A navigation rail, one page per job, and a dark theme that follows your Mac.

Meetings home: the list of recordings and imports, a details panel with the recap — **Meetings**Every call, searchable, with its recap one click away.

A meeting page: transcript, notes and recap — **Meeting**Transcript, your notes and the recap side by side.

The library as a graph of related materials with labels — **Library**Your materials as a graph, labels that assign themselves.

Ask: a question answered with citations to the meeting and a material — **Ask**Answers with citations that open the exact line.

Whisper: a personal inbox of reminders, tasks and ideas said out loud — **Whisper**Say it, and it becomes a reminder, a task or a note.

Requirements: typed statements with the quote each one came from — **Requirements**Typed, quoted, reviewed with the keyboard.

A specification document built from the requirements — **Documents**An SRS or BRD that knows when it is out of date.

The backlog: epics and stories ready for Jira — **Backlog**Stories with acceptance criteria, pushed to Jira.

Settings: speech and language model cards with what fits this machine — **Models**Pick what runs on your machine; Dictum downloads it.

Work

From talk to work. Same afternoon.

Dictum is built for the person who leaves the meeting with the homework. Everything it produces points back to the exact second someone said it.

Recaps that match the meeting.

Brief, bullets or action items, your choice per meeting. The recap rewrites itself as the recording grows, so it is ready when the call ends.

BriefBulletsActions+ your own style

Requirements with evidence.

Functional, non-functional, risks, assumptions and open questions, each with the quote and the timestamp it came from. Review, merge, reject. Nothing invented.

FR · reviewed

Incoming calls show the client card within 2 s, before the agent answers.

26:02 · the card has to come up in under two seconds

Documents that stay current.

An SRS, a BRD, a one-pager. Built from the reviewed requirements and marked out of date the moment a new meeting changes them.

Backlog, straight to Jira.

Stories and acceptance criteria from the requirements, pushed to your board through a browser sign-in. Nothing stored in a keychain.

CRM-41Client card on ringstory

CRM-42Silent listen-instory

Ask the meetings.

"What did the client say about response time?" Answers cite the source; a click opens the line in the transcript.

Under two seconds, measured from ring to card¹; the lookup uses the PBX number².1 Discovery call, 26:02 · 2 Discovery call, 26:41

A library that organises itself.

Drop in the PDFs, the email threads, the old spec. Dictum indexes them on your machine, links what is related and files them under labels you create. New material lands under the right label on its own.

Whisper. A voice inbox for yourself.

Press ⌥space from any app, say "remind me Friday to send Anna the PBX question", and it becomes a reminder. Tasks, notes and ideas land in a personal inbox, kept apart from the project.

REMINDER · FRI 10:00Send Anna the PBX question

Design

Three bars of sound. Three lines of text.

That is the whole idea, and the whole icon. Dictum lives in the Dock and in the menu bar, with a global shortcut for Whisper and a red dot while it records.

Specs

A desktop app, not a tab.

One installer. The first run downloads the models that fit your machine and asks for the two permissions it uses: the microphone and system audio.

14.2macOS

Apple silicon

M1 to M4. 16 GB runs everything locally; 32 GB unlocks the larger Qwen models.

Core Audio tapsMetal

10Windows

64-bit, 10 or 11

A GPU speeds transcription up. It is not required.

CUDACPU

0.4GB

Speech, on disk

The built-in speech model for Russian. The multilingual model is 2.2 GB. Speaker separation with pyannote.

built-in speechmultilingualpyannote

5local LLMs

Language models

Gemma 4 12B, Qwen 3.5 4B and 9B, Qwen 3.6 35B, Qwen 3.8 27B. Claude in the cloud with your own key.

Gemma 4Qwen 3.5Qwen 3.6Qwen 3.8Claude

0accounts

No sign-up, no telemetry

No seat, no server, no usage sent anywhere. Jira and Confluence connect over a browser sign-in when you ask.

JiraConfluenceWordMarkdownZIP

1folder

Everything in one place

Recordings, transcripts, models and settings in one folder you can move, back up or delete. Project files travel between machines.

~/Library/Application Support/Dictum

Your next meeting is on record.

Download for Mac Windows

Free. The source is on GitHub.