Why we built voice dictation that runs on the Apple Neural Engine

When we set out to build Voizer Flow, the fastest path was the obvious one: pipe audio to OpenAI Whisper or Deepgram, get text back, format, paste. That's what nearly every voice tool on the App Store does today.

We didn't take that path. Here's why.

The cloud-first tradeoff

Cloud transcription is great until you think about what you're actually saying out loud all day. For most people, that's:

• Code — including comments, variable names, and the occasional API key
• Client work — names, emails, project details, internal context
• Half-formed ideas — the kind you'd never write down, but voice makes easy

Every cloud dictation tool ships that audio to a third-party server. The privacy policy might say "we don't store it" or "we don't train on it" — but you have to take their word for it. Architecturally, the data left your machine.

Apple Neural Engine to the rescue

Apple Silicon ships with a dedicated Neural Engine optimized for ML inference. As of macOS 14, Apple's Speech framework can run a high-quality speech recognition model entirely on-device, using the Neural Engine instead of the CPU.

Voizer Flow uses this. When you hold the dictation hotkey:

1. Audio is captured into a local buffer
2. The Apple Neural Engine transcribes it in real-time
3. The transcribed text (never the audio) optionally goes to your AI provider for cleanup
4. The waveform is discarded the moment the text lands

No raw voice data ever leaves your Mac. We can't hand it over because we never have it.

The performance surprise

We assumed on-device transcription would be slower than the cloud. It isn't. On an M2 MacBook Air, Voizer transcribes a 10-second clip in roughly 200ms — faster than the round-trip to most cloud APIs.

The Neural Engine is genuinely fast for this kind of workload. We were leaving capability on the table.

What we give up

There are real tradeoffs:

• Intel Macs aren't supported. No Neural Engine means no Voizer Flow.
• The model is smaller than the largest cloud variants. In practice, accuracy is excellent for English; less great for some accents in low-resource languages.
• Custom vocabulary has to be loaded locally instead of hot-swapped on a server.

For our target user — a Mac-native developer or knowledge worker who values privacy — the tradeoffs are worth it. For everyone else, there are excellent cloud-first tools.

The bigger principle

We didn't want to write a privacy policy. We wanted to design a product where the privacy story is structural — where it doesn't depend on us being trustworthy. The Neural Engine made that possible.

If you're on an Apple Silicon Mac, download Voizer Flow and see the difference. If you're not, we're sorry — and we'll see you on Windows eventually.