The Privacy Stack: On-Device, Sandboxed, No Telemetry

Most dictation apps say they are private. Few explain what that actually requires. Real privacy for voice to text is not one setting, it is a stack: on-device speech recognition, a sandboxed app, and no telemetry. Here is how each layer works and why all three matter.

What a privacy stack actually means

When people ask whether Mac dictation is private, they usually picture one question: does my audio get uploaded? That matters, but it is only the first layer. An app can transcribe locally and still leak information through the files it reads or the usage data it reports. True privacy is the combination of three independent guarantees, each closing a different door. If you want the deeper background, we covered the core question in is Mac dictation private, and this piece builds the full stack on top of that.

The three layers are on-device processing, sandboxing, and no telemetry. Think of them as the recording, the storage, and the reporting. A weak link in any one of them undoes the others.

Layer 1: On-device speech recognition

The first layer is where your voice becomes text. With on-device speech recognition, a local model runs the transcription using your Mac's own processor and GPU, so the audio is never sent to a server. BlaBlaType uses local Whisper and Parakeet models optimized for Apple Silicon, and both the recording and the transcript stay on the machine.

This is the layer cloud voice assistants skip. Many web-based speech to text services stream your audio to their servers to transcribe it, which is fast but means your words pass through someone else's infrastructure. OpenAI, for example, documents how its voice features handle audio in its voice mode FAQ. On-device is the opposite approach: the model comes to your voice instead of your voice going to the model. If accuracy is your worry, note that local models are strong now, and the honest way to judge them is how word error rate is actually measured, not marketing numbers.

Layer 2: Sandboxing

On-device transcription protects your audio, but what about the rest of your Mac? That is the job of the second layer. A sandboxed app runs inside boundaries set by macOS: it can only touch the resources it explicitly needs and is granted, rather than roaming your whole disk. For a dictation tool, that means it captures the microphone when you trigger it and types text where your cursor is, without free rein over your documents, photos or messages.

This layer is easy to overlook because it is invisible when it works. It matters most when you think about what a voice tool sits next to: it is active while you write emails, client notes and private messages. Sandboxing keeps that surface small. It also pairs naturally with the difference between dictation and system control, which we unpack in voice commands versus dictation on Mac.

Layer 3: No telemetry

The third layer is the quietest. Telemetry is the background data an app sends about how you use it: how often you dictate, how long, from where, which features you touch. None of that is your transcript, yet together it paints a detailed picture of your habits. An app with no telemetry sends none of it.

This is where a lot of otherwise private tools leak. You can transcribe locally and still report a stream of usage events to an analytics endpoint. Removing telemetry closes that door. Some power users go further and build their own scriptable, local voice control with open tools like Talon. BlaBlaType takes the simpler path for everyday dictation: keep the audio, the transcript and the usage all on the Mac.

Where the AI cleanup fits

Raw speech is messy: filler words, restarts, no punctuation. The best dictation tools clean it up, and here the privacy stack faces its real test, because rewriting usually implies a large model. BlaBlaType runs its AI cleanup on-device using Apple Intelligence, so the polishing step stays inside the stack instead of shipping your draft to a cloud model. If you are curious how that is even possible, we explain how AI rewriting works without the cloud.

Who benefits most from a privacy stack

Everyone gains from a smaller data footprint, but three groups have the clearest reason to insist on all three layers.

Whichever you are, the pitch is the same: pick a tool where all three layers are on by default rather than bolted on. If you are weighing the built-in option first, our Apple Dictation vs BlaBlaType comparison walks through where each one lands, and you can see the plans on pricing.