talat is a local, private transcription app for meetings, dictation, and recordings. Everything you say is transcribed and stored on your own computer. No audio is uploaded, no account is required, and nothing ever trains on your conversations.

Most AI note-takers send your audio to their servers, keep your data there, and cap how long you can retain it. talat runs the whole pipeline on-device: recording, speech recognition, speaker identification, summarisation and storage. It works fully offline, even on an air-gapped machine. Privacy isn't a setting; it's the architecture.

One app, three ways to use it

Record meetings. talat captures your mic and system audio as separate channels, with built-in echo cancellation. It detects Zoom, Teams, Meet, FaceTime and any other conferencing app using your mic and can start recording itself (always, selectively, with a prompt, or manually), then stops shortly after the call ends. A recording notice is always shown; nothing is captured silently. Words appear live as they're spoken, then a high-accuracy pass finalises each segment.

Dictate anywhere. Press a global hotkey and speak into any app on your Mac or PC. The same on-device speech models transcribe what you say, an optional LLM pass tidies punctuation and filler without rewriting your words, and the text is auto-pasted into the focused app. Text replacements, per-app overrides and a searchable dictation history are included.

Import recordings. Drop in WAV, MP3, M4A, MP4, FLAC, OGG or AAC files and talat transcribes them locally through the same pipeline, with the same speaker labelling, custom vocabulary, notes and export targets as a live meeting.

What you get from every meeting

• Real-time, on-device transcription with top accuracy for English, or a multilingual mode spanning dozens of languages

• Speaker identification that learns voices over time and recognises them across future meetings

• AI meeting notes the moment a meeting ends: a summary with inline timestamps, action points with owners, chapters and an auto-generated title

• In-meeting screenshots on a hotkey, anchored to the transcript at the second they were taken and made searchable with on-device text recognition

• Custom vocabulary so names, jargon and acronyms come out spelled correctly

• Inline transcript editing and instant full-text search across every meeting

• Built-in playback with chapter markers; click any transcript line to jump to that moment

• Merge, split, resume and delete meetings, with retention rules you set (7, 30, 90 or 365 days, or forever)

Your choice of LLM

Summaries run on a built-in local model (Qwen3.5-4B) by default. If you'd rather use a cloud model for longer meetings, that's your call: OpenAI, Anthropic, Ollama, your Claude Code or Codex subscription, or any OpenAI-compatible endpoint. You supply the key, it stays in the OS keychain, and you can rewrite the summary prompt to get the notes you actually want.

Fits the tools you already use

• Calendar sync with Apple Calendar and Google Calendar: events become scheduled meetings, attendees are matched to known speakers, and meeting links are detected so recording starts on time

• An MCP server so any MCP-aware assistant (Claude, for example) can read and write your meetings

• Markdown and PDF export with customisable templates, Obsidian export, automatic per-meeting export to a folder, and bulk export of your whole history

• Webhooks that POST or PUT Markdown or JSON to your endpoint when a meeting finishes processing

• Native on macOS (Apple Silicon, macOS 15+) and Windows 10+, with a menu-bar/tray presence, a ⌘K command palette and updates that wait until you're not recording