Sotto

sotto voce — in a quiet voice · そっと — softly, unobtrusively

Fully-local dictation for macOS (Apple Silicon). Hold a hotkey, speak in English or Japanese, release — your words are transcribed with Whisper, lightly cleaned up by a small local LLM (punctuation, filler-word removal), and pasted at the cursor of whatever app you're in.

Everything runs on-device via MLX. No audio or text ever leaves your Mac, and nothing is persisted — audio and transcripts live only in RAM.

• ASR: mlx-community/whisper-large-v3-turbo (EN/JA auto-detect)
• Cleanup: mlx-community/Qwen3.5-4B-4bit (can be toggled off)
• Memory: ~4.5 GB resident while running
• Latency: ~2–3 s for a 10 s utterance after warmup

Setup

From HomeBrew

brew install tomoima525/sotto/sotto
sotto download   # one-time model fetch (~4 GB)
sotto run        # menu bar app

Tap repo: https://github.com/tomoima525/homebrew-sotto

From source

Requires uv and an Apple Silicon Mac.

uv sync
uv run sotto download   # one-time, ~4 GB into the HuggingFace cache

macOS permissions

macOS attributes permissions to the app you launch from — during development that's your terminal (Terminal.app, iTerm2, Ghostty, ...). Grant your terminal all three in System Settings → Privacy & Security:

Permission	Used for	Symptom if missing
Microphone	recording your voice	recordings are silent (rms ≈ 0)
Input Monitoring	the global hold-to-talk hotkey	hotkey silently does nothing
Accessibility	simulated ⌘V to paste at the cursor	text never appears

Microphone prompts automatically on first recording. The other two usually need to be added manually (+ button → select your terminal app). Restart the terminal after granting. If you switch terminal apps, re-grant.

Run the permission doctor to check:

uv run sotto hotkey-test

It prints Accessibility status and DOWN/UP when you press the hotkey; silence on keypress means Input Monitoring is missing.

Usage

uv run sotto run        # menu bar app

Hold Right Option (⌥), speak, release. The cleaned text is pasted into the focused text field and your previous clipboard is restored. (Prefer not to hold the key? Switch Input Mode to Toggle and press once to start, again to stop.)

Input modes

• Hold to talk — hold the hotkey while speaking, release to transcribe.
• Toggle — press once to start, again to stop.
• Streaming (live preview) — press once to start; a floating overlay shows your words as you speak, transcribed phrase-by-phrase by a small fast Whisper model. Press again to stop; the full transcript is cleaned up by the LLM and pasted once into the focused app. The overlay never steals focus. Streaming uses a separate, smaller model for low latency — pick it under Streaming Model: Small (balanced, for daily use) or Base (faster, lower accuracy). Accuracy is lower than the default turbo model, so the final LLM cleanup compensates somewhat.

Menu bar: 🎤 idle · live waveform while recording · ✍️ processing · 🟢 streaming. The recording waveform is driven by your mic level, so a flat line while you speak means the wrong input device is selected. The menu lets you toggle LLM cleanup, choose the input mode, set the language (Universal auto-detects from the audio; force English or Japanese for short utterances that auto-detect gets wrong), pick the microphone (the Microphone submenu shows which device is in use — virtual devices from Loom/Zoom/etc. can silently become the system default), change the hotkey (Right Option / Right Command / F13), and switch the Whisper / Streaming models.

Config lives at ~/Library/Application Support/sotto/config.toml.

Testing each stage

uv run sotto devices                # list input devices, show selected
uv run sotto record --seconds 3     # mic level check
uv run sotto transcribe --seconds 5 # record + Whisper
uv run sotto transcribe --language ja  # force Japanese (auto/en/ja)
uv run sotto stream --seconds 15    # streaming: live segments + final cleaned text
uv run sotto stream --model mlx-community/whisper-base-mlx  # try a faster model
uv run sotto clean "um so I think uh we should ship it"
uv run sotto inject "テスト ✅" --delay 3  # focus a text field within 3s
uv run sotto run --no-menubar       # full pipeline, headless with logs

Limitations

• Doesn't work in secure input fields (passwords) — macOS blocks synthetic events there by design.
• Non-text clipboard contents (images, files) are not restored after pasting.
• The fn key can't be used as the hotkey (not visible to event taps).