Skip to content

CHANGELOG.md

Latest commit

 

History

History
76 lines (64 loc) · 3.23 KB

CHANGELOG.md

File metadata and controls

76 lines (64 loc) · 3.23 KB

Changelog

[0.4.0] - 2026-02-17

Changed

  • Replace Whisper with Kyutai streaming STT: Speech-to-text now uses Kyutai's Moshi-based streaming pipeline instead of batch Whisper inference. Audio is processed as a continuous stream, reducing latency for real-time transcription.
  • Build language model config dynamically from HuggingFace config.json instead of hardcoded Moshi preset
  • Migrate existing Whisper model name settings to Kyutai defaults automatically
  • Simplify pipeline by removing unsafe raw pointer usage

Fixed

  • Frontend crash when checking model status (updated JSON shape for Kyutai backend)

Added

  • Resampling unit tests for Kyutai audio input requirements

[0.3.0] - 2026-02-12

Added

  • Windows support: text output via enigo (keystroke simulation) and clipboard-win (clipboard paste), with platform-gated implementations for Linux and Windows
  • Configurable hold-to-talk hotkey: set hotkey.hold_keys in config.toml to any combination of modifier keys (Control, Super, Alt, Shift). Default: Ctrl+Super
  • GitHub Actions CI for Windows builds alongside Linux
  • Startup grace period (5s) for hotkey listener to prevent false activation from spurious key events during desktop startup
  • Auto-dismiss listening window after 5 seconds of silence
  • Start window hidden on launch to prevent flash of empty window

Fixed

  • Suppress ALSA/JACK stderr noise on Linux
  • Set explicit tray icon ID to prevent duplicates
  • Always auto-hide window on discard

[0.2.0] - 2026-02-10

Added

  • Idle model unloading: Whisper and LLM models are automatically freed from memory after a configurable idle period (default 2 hours). Models transparently reload when the user next triggers dictation, with amber "Reloading models..." indicator. Set gui.idle_unload_seconds in config.toml (0 to disable).
  • Ctrl+Super press-hold-release hotkey for dictation
  • AI cleanup toggle to switch between raw and corrected transcripts
  • Configurable max recording duration (audio.max_recording_seconds)
  • Streaming transcription shows partial results during processing

Fixed

  • Sanitize control characters in text output to prevent terminal issues
  • Black corner artifacts in UI window transparency
  • Waveform display showing only faint blips for quiet speakers

[0.1.0] - 2025-12-06

Added

  • Local speech-to-text dictation using Whisper (CUDA-accelerated)
  • LLM text correction using TinyLlama
  • Tauri overlay window with real-time waveform visualization
  • Click-to-toggle and press-hold recording modes
  • Global hotkey support (Super+H toggle, Ctrl+Super hold-to-record)
  • System tray with show/toggle/copy-last/quit actions
  • Minimize to tray on close/Escape
  • Auto-hide window after successful dictation
  • Click-to-copy transcript text
  • Audio device selection via config
  • Configurable output method (xdotool type or clipboard paste)
  • Configurable typing delay for slow applications
  • Discard recording with Escape key during dictation
  • GNOME shell integration via single-instance --toggle flag
  • Window position persistence across launches (multi-monitor aware)
  • Audio feedback sounds for recording state changes
  • Whisper hallucination and special token filtering
  • Model download manager with verified checksums
  • Comprehensive unit test suite