System / AI & automation systemsArchived

Qiaokeli Dictate

A Chinese-first local dictation tool for KDE: one-hotkey recording, live preview, final correction, and paste-back into the focused field.

Open interactive demo↗Public source↗

01 / EVIDENCE & BOUNDARIES

The repository includes an installer, desktop shortcuts, systemd user services, tray status, a transcription daemon, and troubleshooting documentation.

Tests cover text normalization, domain profiles, device selection, transcript storage, semantic retrieval, personal memory, and command-line interfaces.

Runtime depends on Linux, KDE, and the desktop audio stack; first use downloads Whisper models, and acceleration depends on local drivers and hardware.

This page contains no private datasets, credentials, customer information, or local device access.

02 / HOW IT WORKS

01The Menu hotkey records audio while a small model writes live preview text into the focused field.

02After capture stops, a warm daemon runs the medium model for the final pass and selects CUDA, OpenVINO, or CPU for the available device.

03Normalization, correction memory, and domain profiles repair terminology; the final text replaces the preview and can enter the local SQLite retrieval layer.

03 / WHAT MADE IT USEFUL

One Menu hotkey starts and stops capture, streams preview text while speaking, then replaces it with a full-pass final transcript.

A warm daemon avoids repeated model loading; CUDA and OpenVINO GPU paths are detected automatically with CPU fallback.

Explicit correction memory and domain profiles handle frequent terms; transcript history is stored in SQLite for semantic search and AI-context export.

Green, yellow, and red tray states communicate idle, recording, and transcription; private audio, vocabulary, and history remain local.