Whisper vs Apple speech: what runs locally on Mac

Whisper is OpenAI's open source speech recognition model, released under the MIT license, and it can run locally on a Mac when an app bundles it. Parlo does not use Whisper. It transcribes with Apple's SpeechAnalyzer and SpeechTranscriber, the only engine Parlo ships, entirely on the device. Both approaches can run offline; they use different engines.

By Eric Friedman. Last updated 2026-09-08.

What is Whisper?

Whisper is a speech recognition model that OpenAI released as open source under the MIT license. Its code is public, and a Mac can run it locally after an install, either from the command line or bundled inside another app. Whisper is not a Parlo feature. It is a separate project that other developers, including some dictation apps, choose to run.

What does Apple's speech engine do?

Apple ships SpeechAnalyzer and SpeechTranscriber as part of iOS and macOS. Parlo uses these as its only transcription engine, on both iPhone and Mac. There is no cloud engine behind Parlo and no setting to bring your own API key for a different engine. Transcription runs on the device you are dictating on.

How do the two compare?

Why does an app choose one over the other?

An app that bundles Whisper chooses a model it can install, update, and run on its own schedule. That model does not depend on Apple's release calendar at all.

Parlo takes the other path. It relies on Apple's engine, so transcription improves as Apple improves the operating system. There is nothing to download or manage beyond the operating system itself. The one exception is a one-time on-device speech model download some devices need on first run.

Spokenly, for example, lists Whisper and Parakeet as local models on its free tier, with cloud models on its paid and bring-your-own-key tiers. That is a different tradeoff: more engine choice, in exchange for more to configure.

What does setting up each one actually involve?

Whisper's own README documents installing it from the command line with pip. From there, you run it against an audio file, or wire it into another app's code. That is a developer-facing setup: you, or the app you are using, choose a model size, install dependencies, and manage updates yourself.

Parlo skips all of that. Installing Parlo is the whole setup. There is no model to choose, no package to install separately, and no version of the speech engine to keep up to date. Apple ships and updates SpeechAnalyzer and SpeechTranscriber as part of the operating system, and Parlo simply calls into it.

Does either approach need the internet?

Both can work without a live connection once set up. Whisper needs a one-time local install. Apple's engine is built into iOS and macOS. The underlying on-device speech model may still need a one-time download the first time a device uses it. After that, transcription with either engine happens on the device.

What if Parlo cannot transcribe right away?

If the on-device speech model is missing, Parlo tells you it needs a one-time download. It asks you to connect to the internet and try again. Once that download finishes, transcription runs entirely on the device from then on. If you record and Parlo cannot make out any speech, you see an honest "No speech detected" message. It is never saved as an empty note, as if it were a real dictation.

Could Parlo add Whisper or another engine later?

Parlo ships Apple's SpeechAnalyzer and SpeechTranscriber as its only transcription engine today, with no cloud engine and no bring-your-own-key setting. Adding a second engine, such as a bundled Whisper model, is not part of the current beta. If you want to compare the two yourself, Whisper's own README covers installing it separately from any dictation app.

What happens when Apple updates its speech engine?

Apple ships SpeechAnalyzer and SpeechTranscriber as part of iOS and macOS. An update to either engine arrives the same way any other operating system update does. Parlo does not need to ship an app update to pick up an improvement Apple makes to transcription. The next time you update iOS or macOS, Parlo is already calling into the newer engine. Whisper works the other way around. An improvement to the model only reaches you once the app that bundles it ships a new version with that model included.

What are the honest limits of Parlo's approach?

Tying transcription to Apple's engine means Parlo is English-first today: there is no language picker, and recognition quality follows whatever Apple's engine supports. Whisper, by contrast, ships models trained across many languages, which is one reason an app might choose to bundle it instead. Parlo also has no setting to swap engines if Apple's happens to struggle with a particular accent or vocabulary. The tradeoff for a single, built-in engine is that you cannot pick a different one later without switching apps.

Which should you pick?

If you are choosing a dictation app rather than building one, this comes down to what the app ships, not a setting you toggle. Parlo ships Apple's engine only, on-device processing, English-first, with no cloud fallback and no BYOK option. That keeps setup to installing Parlo itself: no separate model download to manage, no API key to paste in.