dictamac

macOS CLI for transcribing audio files via Apple's SpeechAnalyzer. Agent-shaped: stdout-friendly CLI plus MCP server, both in one binary.

Status: in development. Not yet released.

What it does

Transcribes audio files on macOS using the macOS 26 SpeechAnalyzer / SpeechTranscriber API — the same on-device speech model that powers system features like the Voice Memos transcript view. No model download, no GPU, no Python.

$ dictamac ~/Downloads/meeting.m4a
Hey everyone, thanks for jumping on. So the plan for next quarter is...

JSON output with timestamps:

$ dictamac --json ~/Downloads/meeting.m4a
{
  "version": "1",
  "locale": "en-US",
  "durationSeconds": 184.3,
  "segments": [
    {"startSeconds": 0.0, "endSeconds": 3.2, "text": "Hey everyone, thanks for jumping on.", "confidence": 0.94},
    ...
  ],
  "fullText": "Hey everyone, thanks for jumping on. ..."
}

Voice Memos lookup — find and transcribe a recording the Voice Memos app hasn't transcribed yet:

$ dictamac --voice-memo "yesterday's standup"
$ dictamac --list-voice-memos --since 7d
2026-05-12 14:03  Quick idea
2026-05-12 09:15  Standup recap
2026-05-10 18:22  Walk thoughts

Why it exists

Apple's Voice Memos app transcribes recordings only when it feels like it — there's no public way to force-trigger transcription. dictamac fills that gap, and offers a clean MCP transport so AI agents can transcribe files as part of their workflow.

MCP mode

dictamac --mcp runs an MCP stdio server exposing three tools:

• transcribe_file — transcribe an audio file by path
• transcribe_voice_memo — find and transcribe a Voice Memo by title or date
• list_voice_memos — list recent Voice Memos with metadata

Install

Homebrew (recommended)

brew install jwulff/tap/dictamac

That's it — Homebrew taps jwulff/homebrew-tap, fetches the v0.1.0 source tarball, builds with Swift, ad-hoc signs the binary with the SpeechAnalyzer entitlements, and drops it at /opt/homebrew/bin/dictamac. Build takes about 15 seconds on Apple Silicon.

From source

git clone https://github.com/jwulff/dictamac.git
cd dictamac
make build
make install            # installs to ~/.local/bin by default
make install PREFIX=/usr/local   # or override the prefix

make build ad-hoc signs the release binary with the entitlements SpeechAnalyzer needs at launch. Don't use swift run — it skips the codesign step and the resulting binary will SIGTRAP on first SpeechAnalyzer touch.

The canonical Formula/dictamac.rb lives in this repo so the recipe can be code-reviewed alongside the code it builds; the actively-installed copy is mirrored to jwulff/homebrew-tap on each release.

Requirements

• macOS 26 (Tahoe) or later — uses the SpeechAnalyzer / SpeechTranscriber APIs introduced in macOS 26. Older macOS releases cannot run dictamac.
• Swift 6.x toolchain — Xcode 16 or later, or a standalone Swift toolchain installer. Required whether you install via Homebrew (the formula builds from source) or directly via make build. The formula explicitly checks for Xcode 26+ because that's the toolchain ships the macOS 26 SDK that SpeechAnalyzer lives in.
• Apple Silicon recommended; Intel works but slower.
• Speech Recognition permission — granted on first run via the deep-link printed to stderr.
• Files & Folders permission for Voice Memos library access, needed only for --voice-memo / --list-voice-memos modes.
• Locale model — downloads automatically on first use; needs a working network connection at that moment.

Related

• steno — live capture daemon for macOS speech-to-text, using the same SpeechAnalyzer stack

License

MIT