Skip to content

OSK voice & handwriting (in-process libraries)

Speech and handwriting run inside the Gallery / app process. There is no whisper-cli, arecord, zinnia_character, or PowerShell helper.

We do not train a net from scratch. We load existing pretrained models / OS recognizers:

Feature Preferred library (neural / ML) Built-in fallback
Voice Vosk (libvosk) — Kaldi DNN acoustic model Windows SAPI in-proc recognizer
Handwriting Zinnia (libzinnia) — SVM character model Windows Ink recognizer (language pack)

Platforms: Windows + Linux. macOS: not wired.

Voice (Vosk)

Drop the shared library and a model directory next to the executable, or set:

QWINUI3_VOSK_LIB    path to libvosk.so / libvosk.dll
QWINUI3_VOSK_MODEL  path to a Vosk model folder (e.g. vosk-model-small-cn-0.22)

Also accepted: <appDir>/vosk-model/.

Microphone capture uses Qt Multimedia (QAudioSource) in-process — not arecord.

On Windows, if Vosk is not present, the mic key uses SAPI dictation (install a speech language pack). No PowerShell.

Tap Listen, speak, tap Stop. Text is committed through KeyboardEngine.commitText.

Handwriting (Zinnia)

QWINUI3_ZINNIA_LIB    path to libzinnia.so / zinnia.dll
QWINUI3_ZINNIA_MODEL  path to a Zinnia .model file (e.g. handwriting-zh_CN.model)

Also accepted: <appDir>/handwriting-zh_CN.model.

On Windows, if Zinnia is not present, recognition uses Windows Ink (install a handwriting language pack for Chinese).

Open Settings → Handwriting, draw, tap Recognize, pick a candidate.

Pinyin learning

OskUserLexicon stores local word-frequency in QSettings (QWinUI3/OskUserLexicon). Cleared from Settings → Clear learned words. Boosts candidate order only — no cloud sync.