Skip to content

v1.42.0

Latest

Choose a tag to compare

@goodroot goodroot released this 13 Aug 16:15

This one is for the πŸ‡¨πŸ‡³ πŸ‡―πŸ‡΅ πŸ‡°πŸ‡· folks!

Added

  • append_trailing_space β€” controls the space appended after every transcription: true, false, or "auto" (the default). "auto" checks the last character of the final text and skips the space for Han, Kana, Hangul and full-width punctuation, where a trailing space is a stray character rather than a separator. Everything else β€” Latin, Cyrillic, Greek, and Thai β€” is unchanged. The check runs after post_transcription_hook, so a hook that rewrites the transcription still gets the right answer. (#237, thanks @Leo-Mu)
  • hallucination_markers β€” the list of stock phrases Whisper invents for non-speech audio is now configurable instead of hard-coded. Setting it replaces the built-in list; hyprwhspr config show --all prints the default. Most usefully, "you" can now be removed: it's simultaneously Whisper's most common phantom and a legitimate one-word dictation, and which one matters more depends on how you dictate.

Fixed

  • Word overrides never matched CJK terms β€” multi-character overrides were wrapped in \b…\b, which only matches at a script transition, so {"δ½ ε₯½": "HI"} silently did nothing inside ζˆ‘θ―΄δ½ ε₯½δΈ–η•Œ β€” the common case. Boundaries are now applied per edge and only where one exists to anchor to. This also fixes overrides whose edges aren't letters or digits, like {"c++": "C++"}, which \b…\b could never match. filter_filler_words had the same defect and the same fix.
  • Spaces were welded into CJK transcripts mid-sentence β€” realtime and long-form transcripts joined committed segments with a space, as did the faster-whisper, whisper.cpp and onnx-asr backends when joining Whisper segments. All now join without a space after a CJK character.
  • The ElevenLabs live preview disagreed with what was pasted β€” the mic-OSD pill showed δ½ ε₯½ δΈ–η•Œ while δ½ ε₯½δΈ–η•Œ was injected.
  • The hallucination filter behaved differently in each recording mode β€” continuous mode normalized transcriptions differently from toggle and long-form, so [Silence] and blank_audio were discarded in two modes and pasted in the third. All three now share one implementation.

Contributors

Thanks to @Leo-Mu for the report and diagnosis in #237.

Full Changelog: v1.41.0...v1.42.0