Synthalingua Version 1.2.6 Change Log

[cyberofficial]

Synthalingua Version 1.2.6 Change Log

This update represents a significant evolution for Synthalingua, transforming it from a CLI-only tool into an interactive platform for video transcription. I have also addressed critical stability issues on NVIDIA GPUs and improved the overall efficiency of the batch processing engine.

---

Pick up the portable version on itch!

---

New Feature: Video Translation UI

I'm excited to introduce a full-featured web interface for translating video files. This allows you to visualize the transcription process and edit results in real-time.

• Interactive Player: Upload local video files directly in the browser with standard playback controls.
• Real-time Overlay: Watch translated captions appear over the video as they are processed.
• Waveform Navigation: Visual audio navigation allows for precise seeking.
• Priority Processing: Seeking on the timeline automatically prioritizes the processing of that specific section.
• Subtitle Editor: Review, edit, and export your captions to .SRT format directly from the UI.

How to use:
Run the application with the new flag:

python synthalingua.py --launchui

Performance & Speed

• Model Preloading: Added a --preload argument. This allows you to download and cache Whisper models before running the main application, significantly reducing startup time for subsequent runs.
• Optimized Adaptive Batching: The CPU/GPU allocation logic has been retuned. The system now utilizes the CPU for longer durations (up to 95% load) before forcing GPU-only mode, resulting in better throughput on multi-core systems.

Transcription & Logic Improvements

• Adaptive Rate Limiter: Implemented a rate limiter for HLS streams that detects HTTP 429 errors and applies an adaptive backoff delay to prevent IP bans.
• Internal Repetition Detection: The engine now detects phrase repetitions within a single segment (hallucinations) and filters them out automatically.
• Smart Retry Logic: File processing now identifies low-confidence segments and automatically retries them with a higher-quality model if available.
• FasterWhisper Params: Exposed advanced parameters for the faster-whisper backend, including temperature controls and compression ratio thresholds.

New Utilities

• SRT Repair: New --fixsrt command to reorder existing subtitle files chronologically.
• System Diagnostics: New --bugreport command generates a comprehensive system info file to assist with troubleshooting.
• Browser Cookies: Added support for loading cookies directly from a browser via yt-dlp integration, removing the need to manually export cookie files for authenticated streams.

Bug Fixes

• Critical GPU Fix: Resolved a crash during automatic language detection when using faster-whisper on NVIDIA GPUs.
• PyInstaller Hooks: Fixed a FileNotFoundError related to the transformers library when running from a frozen executable.
• Temp Cleanup: Improved the reliability of temporary file cleanup, specifically for subdirectories created during video processing.

Technical Notes

• CUDA: Updated documentation and build scripts to reference CUDA 12.9.
• Dependencies: Updated requirements.txt to support new UI features (Flask-SocketIO).

---

This discussion was created from the release Synthalingua Version 1.2.6 Change Log.