Upcoming in Synthalingua v1.2.0

⚠️ Note: This is a pre-release changelog. The list of features, improvements, and changes is not final and may be updated before the official release. Some items may be modified, postponed, or withheld to ensure stability and overall product quality.

The next update brings major improvements to stability, performance, and overall user experience. Expect a smoother workflow, smarter transcription, and greater control over your setup.

The Web Interface is under construction and will be set to be released within the 1.2.x version. I have a few more tweaks to make to it. 😫Making the web ui was pain, but hopefully it will prove to be useful.

🚀 New Features

• Force Shutdown System: Prevents server hangs during shutdown using a PID file and watchdog process. Includes a utility script for manual force termination.
• Interactive Stream Selection (--selectsource): Choose your preferred audio format interactively or specify it directly. Perfect for grabbing the best quality stream from Twitch, YouTube, and more.
• Repetition Suppression (--condition_on_previous_text): Reduces redundant or overly similar transcriptions for cleaner results.
• Automatic Blocklist Updates (--auto_blocklist): Frequently blocked terms are automatically added to your blocklist for better filtering.
• Expanded Documentation: Every argument and feature will have detailed usage examples, troubleshooting tips, and clearer guides.
• Improved Model Selection: Smarter prompts to help you choose between English-only and multilingual models based on your available memory.
• Caption Comparison Tool (--makecaptions compare): Generate caption sets using all RAM models at once to compare quality and timing easily.
• Better Console Output: Cleaner logs, new compact mode, and more intuitive progress feedback in the terminal.

⚙️ Improvements

• Streaming Stability: Better error handling, automatic retries, and rate limiting to prevent interruptions or HTTP 429 errors.
• Microphone Input Quality: Upgraded to 48kHz and 24-bit depth for better voice fidelity and less background noise.
• Smarter Subtitles: Improved segment chunking, word-level timestamps, and more accurate timing—especially for fast models like Turbo.
• Language Detection: Improved fallback behavior and support for overriding with a manually selected language.
• Blocklist Handling: Ignores empty lines, adds prompts when missing, and improves error reporting.
• Simplified Web Access: Easier setup and usage examples for using the local web server interface.
• Memory Usage: Improved cleanup routines to free RAM/VRAM when models are no longer needed.

✏️ Tweaks & Changes

• Model Naming: The 12GB model is now correctly labeled as 11GB to reflect its real size.
• Default RAM Model: The default model in setup scripts will shift from 4GB to 3GB for better compatibility on lower-spec systems.
• GUI Wrapper Migration: The graphical interface is now maintained separately and won’t be bundled in the main repo.

🗑️ Deprecated & Removed

• --stream_target_language Deprecated: This argument will be replaced by the upgraded --stream_transcribe command with built-in language support.
• Wrapper Unbundled: The GUI wrapper is no longer included in the core repository.
• Code Cleanup: Removed unused legacy files and deprecated experimental modules.

🐛 Fixes Coming

• FFmpeg stdin Initialization: Fixed crash scenarios when setting up the audio pipe.

Reminder: This changelog reflects planned updates for v1.2.0. Final release content may vary based on further testing and quality assurance. We may hold back certain features to prevent bugs and maintain stability.

Thanks again for your continued support and feedback — more details and a release window will be shared soon!