LocalText2Voice v1.2.0

What's new

• Added automatic SRT and karaoke-style ASS subtitle sidecars from Faster Whisper word timestamps, including chapter-aware exports and narration offsets for podcast mixes.
• Added optional multilingual text normalization with editable SQLite dictionaries for ten languages and independent rules for numbers, ordinals, dates, currencies, percentages, measurements, and Roman numerals.
• Added a second Whisper quality layer that detects unexplained audio after the last source-aligned word, highlights possible trailing artifacts, and contributes to retry candidate selection.
• Added manual audio-tail cutting and optional conservative automatic cleanup. Auto-cut candidates are transcribed again and kept only when the combined review result improves.
• Persisted review metrics and audiobook output paths so subtitles refresh correctly after verification, regeneration, tail correction, rebuilds, UI jobs, and MCP/local-server jobs.
• Made OmniVoice installation reproducible by resolving the tested OmniVoice and PyTorch versions together, using the correct CPU/CUDA indexes and retrying transient Windows cleanup failures.
• Improved Windows uninstall behavior: users can remove downloaded engines, models, runtime dependencies, and voice-gallery data while keeping projects, exports, settings, music, and logs.
• Updated all ten interface languages, documentation, third-party notices, default settings, and the database/settings schemas.

Windows distribution

• Clean installer for Windows 10/11 x64.
• Piper remains the lightweight default engine.
• Optional TTS engines and Faster Whisper remain on-demand downloads.
• Includes the GUI, persistent EngineHost, MCP stdio server, FFmpeg, Piper, bundled voices/music, and the clean embedded Python runtime.
• The installer is currently unsigned.
• SHA-256: 4605bc7e4d590894c28eb59a897d931baade0d3905b9b01aaa00494ec3d2685b

Validation

• 205 tests and 12 subtests passed.
• All ten locale JSON files were validated with 855 matching keys.
• A clean PyInstaller build produced the GUI, EngineHost, MCP server, and embedded Python runtime without optional engine dependencies or local configuration.
• The installer checksum and Windows version metadata were verified.
• An isolated CPU/Piper installation, GUI startup, and silent uninstall completed successfully.