Voice-to-Text That Writes Directly to Your Vault

What it does

Detto is a macOS app for Apple Silicon that captures meetings, takes voice memos, and provides system-wide dictation without sending audio off your machine. It transcribes speech with Parakeet-TDT v3 on the Neural Engine, cleans up the text using Llama 3.2 3B via MLX on the GPU, and writes the result to plain markdown files with YAML frontmatter. The output is built to drop straight into an Obsidian vault or any local agent pipeline that consumes structured text.

The interesting bit

Instead of treating transcription as a dead-end export, Detto reads your existing vault files to surface client context before a call, embeds that metadata into the transcript, and renames the file to keep your knowledge graph organized. Dictation uses a two-phase trick: raw text appears instantly at your cursor while a 3B-parameter model rewrites it in the background, hiding the latency of local inference behind a UX sleight of hand.

Key highlights

• Runs fully offline after a one-time model download; there is no server, account, or analytics SDK.
• Call Capture separates mic audio (You) from system audio (Them) and auto-tags the active conferencing app in the transcript metadata.
• A floating waveform overlay and configurable hold-to-speak hotkey provide system-wide dictation that injects text at the cursor in any app.
• Client Briefing mode scans your vault for type: client frontmatter and surfaces contacts, priorities, and notes before you start recording.
• Audio buffers are discarded after each session; no audio is ever saved to disk.

Caveats

• Requires Apple Silicon and macOS 26 or later, so Intel Macs and older OS versions need not apply.
• The README concedes that small local models “won’t match a frontier cloud model on polish,” so expect occasional refinement quirks.
• Auto-detection supports 25 European languages; support beyond that set is not mentioned.

Verdict

Detto is for Obsidian-dwelling consultants and developers who want their voice to become structured, agent-readable data without touching a cloud API. If you need perfect transcription polish or you are not on Apple Silicon, look elsewhere.