Lalela

From conversation to action, without leaving your Mac

Most AI notetakers ship your audio to a server farm and put a bot in your call. Lalela does the whole pipeline — capture, transcription, speaker separation, even the AI — on the machine in front of you. Here's exactly what happens when you record.

  1. 1

    Record — both sides, no bot

    Hit record (or let auto-detect offer). Lalela captures your microphone and your Mac's system audio directly through macOS APIs, so it hears you and everyone on the call — Zoom, Meet, Teams, phone bridges, or the room itself. Nothing joins the meeting; there is no "Lalela has joined" participant.

  2. 2

    Transcribe — on your Mac

    Whisper runs locally on the Apple Silicon GPU while you talk, biased by your custom dictionary and workspace vocabulary so names, clients and jargon come out right. The audio is discarded after transcription — what remains is text you can read and edit.

  3. 3

    Separate speakers — on-device

    A local diarization model works out who said what and labels every line. Name a voice once ("This is Thandi") and, if you opt in to on-device voiceprints, Lalela recognises her automatically in future meetings. Voiceprints never leave your Mac.

  4. 4

    Summarize — with AI you choose

    One click turns the transcript into a summary in your own prompt style, plus action items with owners. Run it fully offline with a downloaded local model, use the free monthly allowance that comes with signing in, or plug in your own OpenRouter key for any model at cost.

  5. 5

    Use it — search, chat, share, automate

    Everything lands as Markdown files you own, indexed for instant full-text search. Chat with a meeting, a folder, or your whole archive with timestamp-cited answers. Share folders with your team by role, track action items on a shared board, and run multi-step workflows that turn transcripts into minutes, emails or reports.

Under the hood

Whisper on Apple Silicon

OpenAI's open-source Whisper model runs via whisper.cpp on the Metal GPU — fast, accurate, and entirely local. A faster on-device engine option is in the works.

On-device diarization

Speaker segmentation and voice embeddings run with sherpa-onnx models downloaded once to your Mac. Voice recognition across meetings is opt-in and stored only in the local database.

Markdown, not lock-in

Every meeting is a folder of plain files — transcript.md, summary.md, notes.md — in a location you choose. Grep them, sync them, back them up; they're yours.

Local full-text search

A bundled SQLite full-text index makes every word of every meeting searchable instantly, offline, with no search server involved.

Import what you already have

Bring transcripts from other tools (DOCX, PDF, TXT, VTT) or drop in recordings (MP3, M4A, WAV and more) — imported audio runs through the same on-device pipeline, clearly labeled with its provenance.

MCP for your AI tools

A local MCP server exposes your meeting archive to Claude, Cursor and other assistants — on your machine, on your terms.

See it on your own meetings

Free to start — on-device transcription has no caps, and a monthly AI allowance is included when you sign in.

Download for Mac