Whisper apps for Mac: the name tells you nothing

Last updated 11 September 2026

What Whisper actually is

Whisper is a family of speech recognition models published openly by OpenAI. They come in sizes — tiny, base, small, medium, large — trading accuracy against memory and speed. Because the weights are downloadable, Whisper turned high-quality speech recognition from a service you rent into a file you can hold.

That is why so many Mac dictation apps appeared at once. Whisper removed the hardest part of building one.

The three things "uses Whisper" can mean

1. A local Whisper model. The weights sit on your disk, usually run through an optimised implementation such as whisper.cpp. Nothing is transmitted. This is what most people assume and what many apps do.
2. OpenAI's hosted Whisper API. Your audio is uploaded to OpenAI and the text comes back. Convenient, cheap for the developer, and the audio has definitively left your machine.
3. Either, depending on a setting. Increasingly common, and the most misleading, because the default decides what actually happens and the marketing describes whichever option sounds better.

How to tell which you have

1. Look for a model download on first run A local app fetches a few hundred megabytes to a few gigabytes once. An API-based app has nothing to download.
2. Look for a model size setting Choosing between small, medium and large is a local implementation. There is no reason to expose it if a server is doing the work.
3. Run the aeroplane-mode test Network off, dictate. Text means local; an error means an API call.
4. Check whether it asks for an API key If it does, that path is a hosted service — though many apps use a key only for optional extras, so check which.
5. Watch the network if you want certainty Little Snitch or nettop while you dictate. Local apps are quiet apart from licence and update checks.

Whisper is not the only local option

Whisper's openness made it the default, but it is no longer the only serious choice for local dictation. NVIDIA's Parakeet models are notably fast on Apple Silicon and specialise in English; Moonshine and SenseVoice target low-latency streaming. For live dictation — where the text needs to land within a second — those differences are significant in a way they are not for batch transcription.

Vaitly Voice ships Parakeet and Whisper-family models and chooses based on your Mac. The comparison is at Whisper vs Parakeet for local dictation.

Why a dictation app is more than a model

Whisper being open means the recognition is close to a commodity. That is exactly why the model is a poor basis for choosing an app — everyone has access to the same thing. What differs is everything around it:

• Voice activity detection, so the model is not asked to transcribe silence. See what VAD is.
• Getting text into the app you are in reliably, which is harder than it sounds and is where most dictation tools actually fail.
• Custom vocabulary, because the model has never heard of your colleagues.
• Latency management: keeping the model warm without holding a gigabyte of RAM all day.
• What happens when something goes wrong — a history you can replay beats a lost paragraph.