Your Voice Dictation Probably Touches the Cloud. Here Is One That Does Not.
Most voice dictation apps send your audio to the cloud without you knowing. Here's how to verify it \u2014 and an alternative that keeps everything on-device.
The audio you just dictated went somewhere. On most voice dictation apps, it left your computer, hit a server, and came back as text. The round trip is fast enough that it feels local. It isn't.
This matters more for some workflows than others. If you're dictating calendar events and grocery lists, the architecture is probably irrelevant. If you're dictating proprietary code, client requirements, or anything that carries an expectation of confidentiality — it starts to matter a lot.
How Most Tools Work
Standard voice dictation architecture looks like this:
Your microphone captures audio. The app sends that audio to a cloud transcription service — as a compressed file or a real-time stream. The service processes it and returns text. The app puts that text in your active text field.
The cloud service might be Google's Speech-to-Text API. It might be OpenAI's Whisper endpoint. It might be a proprietary service the vendor built. In every case, your audio leaves your device and gets processed on infrastructure you don't control.
Wispr does this. Most AI-enhanced dictation tools do this, because cloud transcription is accurate, cheap to scale, and easy to build on. The vendors aren't hiding it — it's in the privacy policy — but they're also not leading with it.
The latency you feel during dictation — that half-second delay before your words appear — is the round trip. Audio out, text in, via a server somewhere.
What On-Device Transcription Requires
Running transcription locally isn't complicated, but it requires the right hardware and a willingness to do the integration work.
Every Mac with Apple Silicon ships with three compute units: CPU, GPU, and Apple Neural Engine. The ANE is a dedicated chip for neural network inference. It doesn't handle rendering or general-purpose compute. It runs one class of workload: machine learning models.
Parakeet — NVIDIA NeMo's English transcription model, widely regarded as the best open-weight option for English dictation — can be compiled for the ANE using Apple's CoreML framework. The conversion work is significant, which is why most apps skip it and use the GPU instead. GPU tooling is more familiar, PyTorch support is mature, and the result works well enough.
When Parakeet runs on the ANE, the transcription never touches the GPU and never touches the network stack. The chip processing your audio is physically separate from your networking hardware. There is no software path between the ANE and the internet unless you explicitly create one.
VoizerFlow does this work. Your audio stays on the ANE.
How to Verify It Yourself
You don't have to take this on faith. The test takes about thirty seconds.
Open Activity Monitor. Click the Network tab. Start a VoizerFlow dictation session. Watch the outbound data counter.
It doesn't move during transcription.
Compare this to any cloud-based tool. The counter moves every time you speak, even on a fast connection. That's the audio leaving your machine.
This is the difference between verifiable privacy and promised privacy. A privacy policy tells you what the vendor intends to do. Activity Monitor tells you what the software is actually doing.
The AI Cleanup Step
Transcription — audio to raw text — can be fully local. AI cleanup is a separate question.
Most voice dictation apps that offer AI-enhanced output — correcting filler words, adding punctuation, reformatting for the target context — use a language model for this step. Language models are large. Running a capable one on-device produces meaningfully lower quality than calling Claude or GPT-4. So most tools make the API call.
The question is whose API call, under whose agreement, routed through whose infrastructure.
VoizerFlow separates these decisions:
No AI cleanup. Raw transcript, fully offline. Nothing leaves the Mac.
BYOK — Bring Your Own Key. AI cleanup using your own Anthropic or OpenAI API key. The call goes directly from your Mac to the provider you chose, under your API agreement, with your data retention settings. VoizerFlow's servers are not in that path.
Local Qwen 1.5B. AI cleanup running fully on-device. No network call, lower output quality than the API option, complete offline operation.
Each option is a real choice with real trade-offs. The raw transcript is fast and private. The BYOK option gives you the best output quality with one outbound hop to a provider you control. The local model gives you the closest thing to a full privacy guarantee, at the cost of output quality.
Why This Comes Up for Developers Specifically
Consumer voice dictation users dictate notes, messages, and reminders. The risk profile of that audio is low.
Developer workflows are different. The audio a developer dictates in a typical session includes function names, variable naming decisions, architecture discussions, client names embedded in config references, and implementation details for systems that may not be public.
"Implement the rate limiting middleware for the Acme staging environment, same pattern as the auth service, but cap at 200 requests per minute instead of 500" — that sentence contains client identity, environment details, and implementation specifics. If it's being processed by a cloud transcription service, it's on their infrastructure.
For most teams, this is acceptable under their vendor's terms. For some teams, it isn't. For solo developers working on proprietary projects under NDA, the answer often depends on whether they've thought about it.
The architecture of your transcription tool is a decision worth making deliberately, not by default.
What Stays Local, What Doesn't
To make this concrete:
- Audio capture: local on all tools.
- Transcription: local on ANE-based tools (VoizerFlow), cloud on most others.
- AI cleanup: depends on tool and settings. BYOK sends to your chosen provider. Vendor-proxied sends through the vendor's server. Local model stays on-device.
- Usage data and telemetry: varies by vendor, check the privacy policy.
VoizerFlow's claim — audio never leaves your Mac — applies to the transcription step. If you use AI cleanup via BYOK, one outbound call goes to Anthropic or OpenAI. That call contains the text transcript, not the audio. You control the key, the provider, and the agreement.
If you use the local Qwen option, nothing leaves the machine at any point.
VoizerFlow is a macOS voice dictation app for Apple Silicon. Transcription runs on the Apple Neural Engine — audio never leaves your Mac. AI cleanup uses your own API key, sent directly to Anthropic or OpenAI. Or run Qwen 1.5B fully offline. $49.99, one-time.