On-device vs cloud dictation for Mac developers

If you drive a coding agent all day, you type a lot of prompts. Talking is faster than typing, so dictation is the obvious shortcut. The first choice is where the speech to text happens: on a server, or on your own Mac.

Most dictation apps send your audio to a server. A model there transcribes it, and the text comes back. That works. You pay for it with a wait after every sentence, a recurring bill, and a copy of your voice on someone else’s machine.

On-device dictation runs the model on your Mac instead. Akova transcribes with Whisper models on Apple Silicon, so by default your audio never leaves the machine.

Privacy: your prompts describe your code

Prompts for Claude Code, Cursor or Codex are full of repo names, file paths, bugs and plans. That is exactly what you would rather not upload. With on-device transcription there is no upload. Akova keeps audio in memory just long enough to turn it into text.

Cloud transcription exists in Akova only as an opt-in, with your own provider key. If you never turn it on, no audio is sent.

Audio is not the only thing that can leave. Some apps also send your clipboard, selected text or screen text to a cloud model as context. Here is what your dictation app sends besides your voice and how to check it.

Built in voice features in agents are often cloud only. Claude Code’s /voice, for example, streams audio to Anthropic’s servers. Here is Claude Code voice mode vs a Mac dictation app side by side.

Do not take any app’s word for it, ours included. Here is how to check if Mac dictation runs offline with Activity Monitor and two Terminal commands.

Speed: no round trip

Every cloud transcription is a network round trip: upload, queue, process, download. On-device, there is no network in the loop. The prompt lands at your cursor, and you hit Enter.

Offline: dictate on a plane

Because the Whisper model runs locally, dictation works with no connection. Prompt mode, which turns rambling speech into a structured prompt, runs locally too. It uses Apple Intelligence on macOS 26, or a local Ollama model on macOS 14 and 15. A Lifetime licence needs the internet once per Mac to activate, then works offline.

Cost: pay once instead of renting

A cloud service pays for every second of audio it transcribes, so it usually charges a subscription. On-device inference runs on hardware you already own, which is why Akova can be free to start, with a one-time Lifetime licence instead of a monthly plan. See pricing for what each plan includes.

What you give up

Local models need a recent Mac: Akova requires macOS 14 or later on Apple Silicon, and Intel Macs are not supported. Larger Whisper models are more accurate but slower, so you pick the trade off in Settings. For how the model sizes compare, see how Whisper runs offline on a Mac.

Whisper also has one well known quirk: in a long silence it can type a phrase nobody said. Here is why Whisper types “Thank you” when you say nothing, and how to keep it out of your prompts.

Ready to try it? Download Akova free, no account and no card.