Skip to content

WhisperSubTranslate v2.4.2

Choose a tag to compare

View all tags
@github-actions github-actions released this 24 Jun 06:40
· 109 commits to main since this release
v2.4.2
f4a4c8b

What's changed in v2.4.2

Sync accuracy release: a dedicated Faster-Whisper engine repairs subtitle drift, the unreliable SenseVoice timing path is gone, and local translation no longer fails silently or shows misleading API errors.

Added

  • large-v2 Sync model — a separate Faster-Whisper-XXL engine, selectable from the model dropdown, that fixes subtitle timing drift on videos where the normal models slip out of sync. It is most accurate on non-English video (Japanese, Korean, Chinese); English is usually fine with large-v3-turbo. The engine downloads automatically on first use (~4.4GB) with no Python or manual setup. Device behavior is consistent: CPU runs CPU only, GPU runs GPU only, Auto tries GPU first then falls back to CPU.
  • large-v2 Sync Lite model — runs the same Sync model file with int8 compute for lower VRAM (~3GB vs ~4.5GB), for lower-end GPUs. No extra download: Sync and Sync Lite share one model file, so installing or deleting either covers both. Both appear in the model dropdown and Model Manager, localized in all 5 UI languages.

Changed

  • Model Manager wording and VRAM figures — sync model cards spell out what each mode is for and list realistic VRAM (Sync ~4.5GB, Sync Lite ~3GB) from the Faster-Whisper benchmark; large-v3-turbo ~2GB and large-v3 ~4GB corrected from measured runs. Device-lock note clarified so "Auto" no longer reads as GPU-only.
  • README and translated docs — trimmed and corrected (settings file name, build output path, model tables), Contributing split into CONTRIBUTING.md, and all four translations synced.

Fixed

  • Silent untranslated output safety net — if local translation leaves most segments untranslated (the local model crashed mid-run), the job now fails with a clear message instead of saving a subtitle file identical to the source.
  • Misleading "check API key/quota" error on local translation — failures are now reported by method: a local-model failure suggests trying CPU, only online-engine failures mention API keys. Local translation no longer falls back to online engines, respecting the offline choice.
  • Completion sound — no longer hangs waiting on a media event that may never fire; plays reliably on job completion.

Removed

  • SenseVoice timing refinement — the onset-correction path could shift dialogue out of place and is replaced by the large-v2 Sync engine. The bundled SenseVoice model and the sherpa-onnx dependencies are removed, reducing install size.
  • Dead code — unused ETA/formatting helpers, an unused optimal-settings probe, and orphaned split/wrap utilities removed.

Download (Windows portable): WhisperSubTranslate-v2.4.2-win-x64.zip — unzip and run WhisperSubTranslate.exe.

Assets 3
Loading