Dictation on the Mac: cloud or on device
How cloud and on device dictation handle your audio differently, and what I found when I replaced one with the other.
When people pick a dictation app they compare accuracy and price. I did too. The first thing to check is where your voice becomes text.
Where the audio goes
Dictation apps differ mainly in where they process audio. Some record the microphone, upload the audio, and get text back from a server. Some turn the audio into text inside the Mac and never send it anywhere. A few combine the two. From the outside they look identical. Hold a key, talk, let go, text appears. The difference is what happens underneath.
Cloud apps were the default for a long time, for a simple reason. Until recently a Mac could not transcribe Japanese or English fast enough on its own. Uploading audio became so normal that I stopped thinking about it.
What uploading means
Sending your voice means the words you say leave the room the moment you say them. A one line Slack reply, a draft of an email, a note to yourself. Every one of them passes through somebody else's server. The terms of service usually say the audio is not used for training. What a document says and what the app sends are two things you have to check separately.
In September 2026 I opened up one widely used cloud dictation app and looked at what shipped in the bundle. I am not naming it here. The bundle contained a client for a third party transcription service, a Screen Recording permission, access to the Desktop and Documents folders, analytics from two vendors, and an updater that replaces the app without asking. The whole thing was 785 MB. The part that does dictation is a fraction of that.
Why on device became realistic
With macOS 26, Apple introduced a speech API called SpeechAnalyzer. It handles live audio and long recordings, the model runs on the Mac, it streams text while you talk, it covers Japanese and English, and there is no fee per use.
The other option is Whisper, the model OpenAI released. Through the whisper.cpp implementation it runs on any Apple Silicon Mac. Yubi supports both engines, and neither sends audio off the machine.
How to tell which kind you have
Permissions tell most of the story. Open System Settings, then Privacy and Security, and look at what the app asked for. Yubi uses two things: the microphone, and Accessibility so it can type into the app in front of you. An app that also wants Screen Recording or your folders is doing something beyond dictation.
The other signal is an account. A sign up screen means the app creates an account for you. That alone does not show what usage data it collects, but software that finishes its work inside your Mac has no reason to make you log in.
Does accuracy suffer
In my everyday use, SpeechAnalyzer's Japanese felt close to the cloud service I used before. Names and company names trip up every engine, so Yubi has a box where you list the words you use a lot. Whisper transcribes after you stop talking, so you wait a moment, and on my Mac it holds together better in noisy rooms.
Keeping the audio on my Mac mattered more to me than small differences in recognition, so I built Yubi.