On-device meeting transcription, explained (no cloud upload)
Last updated: July 2026
For years, turning a meeting into text meant sending the audio to a server. That is finally optional. On-device transcription runs the speech-to-text model on your own computer, so the audio never leaves it — and on a modern Apple Silicon Mac it is fast enough for real work. This explains what on-device transcription is, the models that made it practical, what you gain, and where it still falls short.
What “on-device” actually means
On-device — sometimes called local or edge — means the automatic speech recognition (ASR) model runs on the hardware in front of you rather than in a data center. You press record, the audio is fed to a model loaded into your Mac’s memory, and text comes back, all without a network request.
The consequences are direct: no upload, no account required for the transcription itself, no per-seat cloud bill, and it keeps working on a plane with the Wi-Fi off. The privacy story is the honest headline — audio that is never sent cannot be stored, subpoenaed, or used to train someone’s model.
The models that made it practical
On-device ASR went from clumsy to genuinely good in a short span, driven by three families of models:
- WhisperKit — an open-source Swift package from Argmax that runs OpenAI’s Whisper models on Apple Silicon via Core ML, scheduling the work across the Neural Engine, GPU and CPU. It has been available since early 2024 and, as of 2026, ships alongside on-device speaker diarization and text-to-speech in one MIT-licensed package.
- Apple SpeechAnalyzer — Apple’s own framework, introduced at WWDC 2025 and built into macOS 26 and iOS 26. Its SpeechTranscriber model was trained for long-form audio — meetings, not just short dictation — and Apple reports it running roughly twice as fast as Whisper Large V3 Turbo on comparable tasks, with nothing leaving the device.
- Parakeet — NVIDIA’s compact ASR model (the 0.6-billion-parameter parakeet-tdt) that tops open transcription leaderboards on accuracy. A community MLX port runs it on Apple Silicon, where it can transcribe an hour of audio in under a minute, and the multilingual v3 covers 25 European languages.
The point is not which one wins. It is that a laptop can now transcribe a meeting locally, quickly, at an accuracy that used to require the cloud.
What you gain
The obvious win is privacy: for anyone under an NDA or handling a client’s confidential discussion, “the audio never left my Mac” is a far stronger position than “the vendor says they delete it.” But there are quieter benefits too. There is no seat-based subscription for the transcription — the model is a file on your disk, not a metered API. There is no outage to wait out and no rate limit. And there is no ambiguity in a terms-of-service document about whether your meetings might be retained or used to improve someone’s product.
Where on-device still has limits
Being honest about the trade-offs matters. On-device transcription is not magic:
- Speaker separation is the hard part. Turning a transcript into “who said what” — diarization — is genuinely difficult, and even good on-device diarization can stumble on crosstalk, similar voices or a poor microphone. It has improved a lot, but do not expect flawless labels on a chaotic six-person call.
- Models cost disk and memory. A capable model is hundreds of megabytes to a couple of gigabytes, and it uses real RAM while running. On an older or low-memory machine that is a genuine constraint.
- Language coverage varies by engine. Whisper is broad; Apple’s and Parakeet’s coverage is strong but different. The right engine depends on the languages you actually speak.
For plenty of teams a cloud tool with a polished shared workspace is still the better fit, and that is a fair choice — folo vs Granola lays out where a synced cloud notepad wins.
On-device by default, cloud only if you ask
folo is built around this idea. It transcribes on your Mac using exactly these engines — WhisperKit, Apple SpeechAnalyzer, or Parakeet — with on-device speaker recognition that remembers a named voice across future meetings. A cloud transcription option exists, but it is opt-in; the default is local, and the default is the point. If keeping meeting audio on your own machine matters to you, on-device transcription is no longer a compromise — it is just how it should work.
Questions, answered
What is on-device meeting transcription?
It is speech-to-text where the ASR model runs on your own computer instead of a server, so the audio is never uploaded. On an Apple Silicon Mac it is fast enough for real-time meeting transcription.
Is on-device transcription as accurate as the cloud?
For the audio itself, modern on-device models — Whisper, Apple SpeechAnalyzer, Parakeet — are competitive with cloud ASR. The weaker spot is speaker diarization (labelling who said what), which is hard everywhere but especially on messy multi-speaker calls.
Which on-device transcription models can a Mac run?
WhisperKit (Whisper on Apple Silicon via Core ML), Apple’s built-in SpeechAnalyzer on macOS 26, and NVIDIA’s Parakeet via an MLX port. Each has different language coverage and speed, so the best engine depends on what you record.
Does on-device transcription need a subscription?
Not for the transcription model itself — it is a file on your disk, not a metered cloud API. Tools built around on-device transcription are often a one-time purchase rather than a per-seat subscription.