Every dictation app makes the same promise: speak naturally, get clean text. The difference between them is a question almost nobody asks: where does the work happen?

For most dictation tools, the answer is: on the vendor's servers. Your microphone audio is captured, uploaded, transcribed in their cloud, cleaned up by their models, and the text is sent back down. That architecture is easy to build and easy to scale, and it means your voice, the most personally identifiable signal you produce, routinely leaves your machine.

We built ShoutFlow on the opposite bet: your Mac is already fast enough to do the whole job.

The two jobs of dictation

Dictation is really two distinct problems:

Transcription: turning audio into raw words.
Clean-up: turning raw words into writing: stripping the "umms", resolving "wait, no" self-corrections, fixing punctuation, matching the tone of the app you're in.

ShoutFlow treats these as independent stages with independently selectable engines. In local mode, both run on-device:

Transcription runs on WhisperKit, which executes Whisper-family speech models on the Apple Neural Engine via Core ML. The Neural Engine is purpose-built silicon that sits mostly idle in day-to-day use. Dictation is exactly what it's for.
Clean-up runs on MLX, Apple's array framework for machine learning on Apple Silicon, which executes a small language model on the GPU.

Neither stage needs the network. Flip on Airplane Mode after the one-time model download and ShoutFlow keeps working. That's not a marketing line, it's a supported configuration we test.

"Private by policy" vs "private by architecture"

Cloud dictation vendors publish privacy policies, and many of them are sincere. But a policy is a promise about behaviour: we receive your audio, and we promise to handle it well. Promises can change with an acquisition, a subpoena, a breach, or a product pivot.

An architecture is different. In ShoutFlow's local mode there is nothing to promise about, because there is nothing to receive. Our servers physically cannot leak, retain, train on, or hand over your audio; it never arrives. The only thing the app ever tells us is whether your license is valid, and that check carries no audio, no text, and no usage data.

The strongest privacy guarantee is the one that doesn't depend on anyone keeping it.

What about quality?

The honest trade-off: the largest cloud models are stronger than the small models that fit on a laptop. For dictation clean-up, though, the job is narrow (fix punctuation, remove filler, apply a tone), and small local models handle it remarkably well. You also choose the model size yourself: lighter models are quick on older machines, heavier ones are more meticulous.

And when you genuinely want a frontier model doing your clean-up, ShoutFlow doesn't force a choice between privacy and power: bring your own key and your Mac talks directly to the provider you trust, with no middleman in the path.

Why this matters more every year

Speech models keep shrinking while Apple Silicon keeps getting faster. Each generation of chip makes the "upload your voice to someone else's computer" architecture less necessary. We think local-first is simply where this product category ends up. We just didn't want to wait.

ShoutFlow is $25 once, and the free trial needs no account and no card. Try the Airplane-Mode test yourself.