typelessless

A voice-dictation tool you own end-to-end — a self-hosted alternative to Typeless for Windows, built for people who speak 中文 and English mixed together. Press a key, talk (switching languages mid-sentence), press again — cleaned text is inserted into whatever field has focus. No subscription, your own API keys, your own prompts.

Right Alt (toggle) → mic (16 kHz) → Soniox STT  (语言提示 en+zh, 词汇偏置)
   → mode router (auto by focused app) → Claude cleanup → paste into focused field

Features

• Real 中/英 code-switching via Soniox realtime STT — mid-sentence language switching, no manual toggling, with custom-vocabulary bias.
• LLM cleanup via Claude, with a global "core" prompt plus per-mode prompts. Four default modes, escalating in aggressiveness: chatting (verbatim+) · working (remove fillers/false starts) · cleaning (reorder/merge redundancy) · polish (cross-sentence merge + rewrite).
• Auto mode selection by focused window — e.g. focus VS Code → working; configurable app→mode rules, else a default mode.
• Settings web UI (tray → Settings) to create/edit modes, the global prompt, the vocabulary, the default mode, and the routing rules — no file editing.
• History web UI (tray → Show history) — every dictation is logged with its audio; retry re-transcribes a saved recording and copies the result. The raw audio is saved before any network call, so a failed request never loses a recording.
• Floating waveform indicator at the bottom-center of the monitor under your cursor (red = recording, amber = processing).
• Mutes system audio while recording (restored on stop).
• Push-to-talk toggle on Right Alt (fully suppressed so it doesn't leak to other apps); walkie-talkie hold mode also available.
• Login autostart toggle; runs as a tray app with no console.
• Provider-facing code sits behind interfaces — swap Soniox or Claude (or a local model) without touching the app.

Bring your own keys

You supply your own Soniox key (STT) and, for cleanup, an Anthropic key. They live in config.toml, which is gitignored — it is never committed. You can also set $SONIOX_API_KEY / $ANTHROPIC_API_KEY instead and leave the file blank.

Rough cost: Soniox ≈ $0.10–0.12 / hour of audio; Claude Haiku cleanup is a fraction of a cent per utterance. A few dollars a month in normal use.

Install (Windows, single .exe)

Windows with Python 3.11+:

git clone https://github.com/EricWay1024/typelessless
cd typelessless
./build.ps1

build.ps1 creates a venv, installs everything, and produces dist\typelessless.exe plus a dist\config.toml. Then:

1. Edit dist\config.toml — add your Soniox key (and Anthropic key for cleanup).
2. Double-click dist\typelessless.exe — a tray icon appears.
3. Press Right Alt, speak, press Right Alt again → text is inserted.
4. Tray → Settings to customize modes/prompts/vocab/rules; Show history to review, play back, and retry past dictations; Start on login to autostart.

Run from source (dev)

./run.ps1     # venv + install + config.toml + run, in one step

Verify the STT+cleanup pipeline against an audio file on any OS (no mic needed):

pip install websockets anthropic
cp config.example.toml config.toml     # add keys
python -m typelessless filetest sample.wav working
python -m typelessless check            # verify all components load

Configuration

• config.toml (system settings, hand-edited): API keys, STT model, language_hints, hotkey, injection method, mute_while_recording. See config.example.toml.
• Modes, prompts, global prompt, vocabulary, and app→mode rules are managed in the Settings web UI (tray → Settings) and stored in %APPDATA%\typelessless\settings.json. On first run this is seeded from the built-in defaults (src/typelessless/settings.py).

Logs and recordings live under %APPDATA%\typelessless\ (dictations.log, history.json, recordings\).

Layout

src/typelessless/
  config.py / settings.py   system config (toml) + editable settings (json)
  modes.py                  Mode dataclass
  audio.py                  mic capture (sounddevice)
  stt/{base,soniox}.py      STT interface + Soniox realtime WebSocket client
  cleanup/{base,claude,passthrough}.py
  inject.py                 Windows paste / unicode-type
  hotkey.py                 global hotkey — toggle or hold, key-suppressing (pynput)
  overlay.py                floating waveform indicator (Win32/ctypes)
  sysmute.py                mute system audio (Core Audio COM via ctypes)
  foreground.py             focused-window detection for mode routing
  history.py / webui.py     recording store + local history/settings web UI
  dictlog.py / startup.py   dictation log + login autostart
  tray.py / app.py          tray menu + orchestrator
  filetest.py / selfcheck.py
typelessless.spec           PyInstaller build → dist/typelessless.exe
build.ps1 / run.ps1

Prebuilt binaries & licenses

This project is MIT-licensed (see LICENSE). It depends on pynput and pystray (both LGPL-3.0), used as ordinary pip packages — this does not affect the license of this code. If you distribute a prebuilt typelessless.exe (PyInstaller-bundled), note that LGPL asks that recipients be able to replace those libraries; the simplest way to satisfy that is to point users at this source and build.ps1 so they can rebuild. Distributing the source has no such caveat. Other dependencies (anthropic, websockets, sounddevice, pillow, pyperclip) are MIT/BSD.

Status

Windows desktop app (working). An Android IME reusing the same Soniox + cleanup contract is a possible next phase.