A dictation tool that is always listening has already made the expensive trade.
Sotto voce — in a soft voice. Sotto is for people who write in two languages all day and would rather talk than type: press a key, speak, press it again. Audio is transcribed on your own machine with faster-whisper, cleaned up, and pasted where the cursor already is. F13 for 中文, F15 for English.
▶ Download for macOS or Windows — .dmg for Apple Silicon, portable .exe for Windows 10/11.
1. An explicit key press, not an open microphone — because the two failure modes cost different amounts.
Always-listening buys about one second of convenience. It costs a live mic and a class of failure that writes text into whatever window you happen to be in. Sotto's audio stream opens on a press and closes on a press, and the edges are handled rather than assumed: takes under min_duration (0.3 s) are discarded as accidental taps, and a forgotten stop is capped at max_duration (300 s) and transcribed rather than buffered forever. Modifier keys are rejected outright as dictation keys — a stray ⌥ while typing would silently start recording, the exact bug this design exists to avoid.
It is press-to-toggle rather than hold-to-talk because holding a key through a two-minute thought is a physical tax, and because macOS repeats key-down events while a key is held.
2. One key per language beats a language picker, because a picker is a step you take before you know you need it.
Whisper can auto-detect, but detection drifts mid-sentence — the dominant failure when you switch languages often rather than occasionally. Sotto hands the model the language up front, keyed to the hotkey you hit: {"f13": "zh", "f15": "en"}. Pressing the other key while recording ends the take instead of switching mid-sentence. "auto" is available per key if you want it.
3. Polish is a spectrum with a network boundary in the middle, so it is a mode, not a checkbox.
| Mode | What runs | Leaves the machine |
|---|---|---|
off |
Raw Whisper output | No |
local |
Regex cleanup — a narrow filler list (um/uh/erm/hmm, 嗯/呃/唔), repeated-word collapse, spacing, punctuation |
No |
ai |
Claude (claude-haiku-4-5) applies mid-sentence self-corrections, formats dictated lists, fixes tone |
Yes |
auto |
ai if ANTHROPIC_API_KEY is set, else local |
Only if you set a key |
The local tier removes only tokens that are essentially never meaningful, so it cannot change what you said; the AI tier can do more, and is therefore opt-in. If the API call throws, it falls back to local_polish rather than erroring — a dictation tool that loses your words once does not get used twice.
4. The personal dictionary is worth hand-maintaining, because transcription error is not uniformly distributed.
Whisper is fine on common words and unreliable on exactly the words carrying the meaning: names, product names, jargon. ~/.sotto/dictionary.txt attacks that from both sides — a bare term (Kubernetes) goes to Whisper as its initial_prompt (first 40 entries), biasing recognition before the mistake; a rule (cloudy tan -> Chloe Tan) rewrites the output after it. Twenty lines you maintain once beat a larger model on your own vocabulary.
press F13 ──▶ record ──▶ press F13 ──▶ Whisper (local) ──▶ polish ──▶ ⌘V insert
The menu bar carries model size (tiny…large-v3), polish mode, translate and the dictionary; a capsule shows which language is recording; the previous clipboard is restored 0.6 s after the paste. Translate to English uses Whisper's own translation task — offline, and only into English.
macOS (Apple Silicon): open the .dmg, drag Sotto to Applications, then right-click → Open → Open the first time (not notarized — no paid developer certificate behind this project). Grant Microphone and Accessibility in System Settings › Privacy & Security, then quit and reopen once; macOS only applies a grant to newly started processes. That is the whole permission surface — no screen recording, no camera, no cloud.
Windows 10/11 (beta): run Sotto.exe — portable, nothing to install. SmartScreen → More info → Run anyway; first start unpacks for ~20–30 s.
From source — needs uv; Python 3.11–3.14 is fetched for you:
git clone https://github.com/chloe4ai/sotto.git && cd sotto
uv sync # add --extra ai for Claude polishing
uv run sotto
uv run sotto --doctor # checks permissions and setupFirst run downloads the Whisper base model (~150 MB, once); offline afterwards. ~/.sotto/config.json holds hotkeys, model, polish mode, task and durations — any pynput key name plus a Whisper language code.
Nothing happens on a key press: almost always Accessibility (plus Input Monitoring on some macOS versions); run --doctor. macOS silently revokes the grant when your terminal updates — toggle it off and on. Mic blocked: the permission belongs to the app you launched from, and if you once clicked Don't Allow, macOS never asks again (tccutil reset Microphone <bundle-id>). F13–F20 don't exist on laptop keyboards — pick other keys or remap with Karabiner-Elements.
- No accuracy number. There is no WER harness against a held-out set of my own speech, so every claim here is about the design, not the recognition quality.
- The dictionary biases, it does not guarantee.
initial_promptis a hint, and only the first 40 terms are sent. - One language per take — deliberately, per decision 2, but code-switched sentences do come out worse than either language alone.
- Unsigned on both platforms. Read the code or build from source if that matters.
- A WER harness over a fixed set of my own recordings, per model size and per language, so "switch to
small" becomes a number and the dictionary's contribution is measured rather than assumed. - Speak-to-edit: select text, speak an instruction. Measure how often the edit is accepted unmodified — an edit you fix by hand is slower than typing.
- Hotkeys settable from the menu bar, so the commonest setup failure stops requiring a JSON file.
MIT
