Sprachacus.

Common dictation tools upload the recording to a server and charge a monthly fee. For notes that is merely annoying; for client calls and internal meetings it is simply not an option. At the same time typing remains slower than speaking — the need does not go away.

Sprachacus uses the SpeechAnalyzer framework introduced with macOS 26 and transcribes locally while you speak:

• Dictation via the right Option key, text inserted at the cursor — in any application
• Assist mode: copy a text, speak an instruction, review the finished reply
• Meeting transcription with separate tracks for microphone and system audio, followed by a structured summary
• Layered clean-up: Apple Intelligence locally, then the Claude CLI, otherwise raw text — insertion never fails
• Swift 6, no dependencies, open source under GPL-3.0

In Assist mode the text is foreign — a copied email may contain instructions aimed at the AI. In testing, the local Apple model followed such an embedded instruction in three out of three runs, even with a hardened protective prompt. Claude Haiku resisted. That is why dictation runs locally while Assist and meetings go through the CLI. Measurements like this decide our architecture — not assumptions about what ought to work.

0 €

transcription runs locally and costs nothing

A tool that does not exist yet?