macOS 13+ · APPLE SILICON

Your voice never
leaves your Mac.

Soufflé dictates into any app, transcribes meetings that tell your voice from everyone else's, and writes the summary, all of it on-device. No cloud, no account, no API key.

Download for macOS

Free and open source

Soufflé · Live meeting ReadyKyutai STT 1B
Live meeting
12:40 Stop

Live transcript

System audio active

Double-click a paragraph to fix a misheard word. Later occurrences are corrected automatically.

Notes

Saved
Angela owns the client comms. The dry run passed on staging. Tyrell still has to sign off on the disclosure.

A live meeting transcript, separating you from everyone else. No virtual audio device, nothing uploaded.

Meetings

Meetings that know who is talking.

System-audio capture separates Me from Them in the live transcript, with no virtual audio device to install. Recording stays hands-off: it offers to start when a calendar meeting begins, notices when the meeting seems over, and picks back up where it left off when your Mac wakes.

Optional Opus audio with a retention policy, replayable by clicking the transcript
Recovers or salvages the session if the engine stalls or the mic disappears
Everything grouped by day, with full-text search across every word
Soufflé ReadyKyutai STT 1B
Dictate Press⌘⇧Space
Meeting Live transcript · captures system audio
Search transcripts and meetings

Coming up today

Product syncNow 4 10:00–10:45
1:1 with Angela 14:30–15:00

Today

Product sync 32:1010:00
Reply to the security audit thread. I think we should hold the current access tier until… 00:4109:12
Design review 18:4408:30
Soufflé · Dictating ReadyKyutai STT 1B
Dictating
0:00 Stop

Auto-pastes into your last app when you stop
Dictation

Speak a paragraph, get a paragraph.

With the default model the text streams in while you are still talking, punctuated and capitalised. A local pass can tidy the phrasing before you use it, and that same model handles French and English without you switching anything.

Text appears as you speak, not once you stop
Press once to start and stop, or hold to talk
Fix a misheard name by hand once, and it stays fixed
A local LLM pass can polish the wording, with editable prompts
French and English on the same model, no switching
Anywhere you type

The same dictation, in every app you use.

Soufflé is not a window you type into. Press the shortcut and a small overlay appears above whatever you are in, then the text lands in the field your cursor was already in. Chat, mail, an editor, a terminal: the overlay does not care which, and it never turns up in a screen recording.

The text lands where your cursor already was, with no copy and paste
Terminals and secure fields can get simulated keystrokes instead
Optional start and stop sounds, so you know the shortcut landed
E Corp Chat
#release-4-2
9 Ship checklist and blockers for 4.2
A
Angela Moss09:58

The client is asking when we can announce the disclosure. What do I tell them?

T
Tyrell Wellick10:02

Depends on the key rotation, no? I'd rather not commit to a date we then move.

Message #release-4-2
ShortcutDictatingReformulatingPasted

Press the shortcut anywhere, talk, and the tidied text lands where your cursor already was.

Summaries

The summary is written on your machine too.

Decisions, action items with owners, and the questions nobody answered, extracted alongside the summary by Apple Intelligence, or by a local Ollama when your Mac does not offer it. Transcription itself needs neither.

Editable notes and participants next to the transcript
No setup at all when Apple Intelligence is available
Soufflé · Product sync ReadyKyutai STT 1B

Outcomes

Decisions

Hold the deploy until the key rotation landsPentest review moves to Thursday

Action items

Rotate the encryption keys

Owner: Elliot Alderson

Tell Darlene the timeline moved by two days

Owner: Angela

Re-run the security dry run on staging

Owner: Elliot Alderson

Book the follow-up for next Tuesday

Owner: Tyrell

Open questions

Do we need a fallback if the rotation runs long?Who signs off on the disclosure?
Speech models

Pick the trade-off you want.

All four run locally, downloaded on first use. The two Kyutai models stream text as you speak; the other two transcribe once you stop.

Kyutai STT 1B

Default

French and English, streaming as you talk. The one most people should stay on.

2.4 GBStreamingMetal · CandleFR + EN

Kyutai STT 2.6B

English only, streaming, noticeably higher quality if you have the disk and the patience.

5.6 GBStreamingMetal · CandleEN

Whisper Large V3 Turbo

Multilingual and familiar, through whisper.cpp on the GPU. Transcribes once you stop.

1.6 GBOn stopMetal · whisper.cppMultilingual

Parakeet TDT 0.6B v3

Smallest

25 languages with punctuation and capitalization, fast on the CPU alone. Transcribes once you stop.

670 MB int8On stopCPU · ONNX25 languages
Privacy

Private by design, not by promise.

Transcription, summaries and audio all happen on your Mac. There is no server to send them to.

Nothing is uploaded

Your audio, transcripts, notes and summaries stay on your Mac, in one local database you can export or delete whenever you like.

Works offline

Once the speech model is on disk, transcription keeps working with the Wi-Fi off, on a plane, anywhere.

No account, ever

No sign-up, no subscription, no API key to paste in. You download it and it works.

The app makes exactly three outbound requests. See what they are, and why.
Your data

Everything stays yours to take with you.

No proprietary format, no export fee, nothing holding you in.

Export anything

One meeting or the whole archive, as ordinary files you can open in anything.

Delete it just as easily

Settings clears the lot in one action, recorded audio included.

Plug in your own tools

Your transcripts can be searched and read by tools you already use, read-only and on your machine.

Export formats, the MCP server and the command line are in the docs.

What you need

Mac Apple Silicon. There is no Intel build.
macOS 13+ for dictation, 14.4+ to capture
the other participants
Disk 670 MB to 5.6 GB, for the model you pick
Summaries Apple Intelligence (macOS 26+)
or a local Ollama

Permissions, asked when needed

Never up front, and Settings shows the state of each with a link straight to the right System Settings pane.

Microphone Records your voice
System Audio Captures what the others say
Accessibility Pastes into the app you were using
Calendar Lists today's events, optional

Open source, actively developed. Every release is signed and notarized, and the engineering is covered by a large test suite. Macs differ enormously in audio hardware and routing, so if something does not behave on yours, please open an issue with the details.

Try it on your next meeting.

Download the .dmg, pick a model once, and the rest happens on your Mac.

Download for macOS Source on GitHub