A Look at Beta 1.1.1

[cyberofficial]

Hello everyone!

As a solo developer, I've been hard at work on a massive architectural update for Synthalingua. This isn't just an incremental patch; it's a foundational rebuild of the core engine designed to be faster, more stable, and packed with powerful new features.

Before I finalize everything for a stable (1.2.0) release, I'm going to publish a 1.1.1 beta build which will help me test in live environments.

How to Test It Now (For Advanced Users)

This is a preview for those comfortable with running from source code. A polished, portable build for all users will be released after this testing phase is complete.

If you'd like to jump in and test it today, you have two options:

• Check out the Pull Request: Pull Request #159
• Download the branch directly: refactor branch

---

What's New: A Look at the Next Generation of Synthalingua

Major New Features

• Crystal-Clear Transcriptions with AI Vocal Isolation: A new feature (--isolate_vocals) that uses the Demucs AI model to intelligently separate spoken words from background music, noise, and other interference. This allows you to get clean, high-quality transcriptions from audio that was previously unusable.
• Interactive Stream Selection with Audio Preview: The new --selectsource flag allows you to see all available audio tracks from a live stream and listen to a short preview before starting a full transcription. This eliminates the guesswork and ensures you're always transcribing the correct source.
• More Natural Live Microphone Transcription: A new context-aware mode (--paddedaudio) allows the AI to remember what you just said, using it as a reference for what you're about to say next. The result is far more fluid and accurate live transcriptions with fewer awkwardly cut-off sentences.
• Advanced Subtitle Generation: The subtitle generator now features a --makecaptions compare mode to generate subtitles using every available AI model, so you can easily choose the best result. It also now supports --word_timestamps for creating karaoke-style subtitles.

Fixes, Improvements, and Stability Upgrades

• Modern, Readable Interface: The command-line output has been completely redesigned with clean formatting, colors, and structured boxes to make it much easier to read.
• Smarter Repetition Filtering: The application now detects and blocks when the AI gets stuck in a repetitive loop. An experimental --auto_blocklist feature can even add these looping phrases to your ignore list automatically.
• Professional Remote Microphone Tool: The remote_microphone.py script has been transformed into a powerful, standalone utility for setting up remote audio streaming over your local network.
• More Resilient Streaming: If a streaming service temporarily rate-limits the connection, the application now intelligently waits and retries instead of crashing, making long-running streams far more stable.
• Automatic Memory Cleanup: The tool now automatically unloads the AI model from memory (RAM/VRAM) after processing a file, preventing slowdowns and out-of-memory errors when working with multiple files.
• Easier Setup: Getting started is simpler. The app now automatically searches common locations for your cookies file and offers to create a template ignorelist file if one doesn't exist.

---

For Advanced Users: Technical Deep Dive

1. Core Architecture and Performance

Engine Overhaul: Asynchronous Queue-Based Processing

• Implementation: The previous linear "record-process-repeat" loop in TranscriptionCore has been replaced with a multi-threaded, producer-consumer model using threading and queue.Queue. The main thread captures audio data (producer) and places file paths into a queue. A separate background thread (consumer) retrieves paths from the queue and performs the computationally expensive Whisper model inference.
• Technical Impact: This decouples audio I/O from model computation, preventing dropped audio during high CPU/GPU load and significantly improving real-time responsiveness. It effectively turns the transcription process into a parallel pipeline.

Memory Management in sub_gen.py

• Implementation: A new unload_model() function has been introduced. After each transcription task in run_sub_gen, this function is called to del the model object, force gc.collect(), and, if applicable, run torch.cuda.empty_cache().
• Technical Impact: This aggressively frees VRAM and system RAM, which is critical for the --makecaptions compare mode. It allows the script to iterate through multiple large models without causing memory fragmentation or out-of-memory errors.

2. Stability and Error Handling

Adaptive Rate Limiting for HLS Streams

• Implementation: A new rate_limiter.py module provides a global RateLimiter class. The HLS segment downloader in stream_transcription_module.py now catches HTTP 429 "Too Many Requests" errors and other network exceptions. When a 429 error occurs, it uses the rate limiter to apply an adaptive backoff delay, which increases with subsequent failures and resets upon success.
• Technical Impact: This makes the HLS stream ingest far more resilient to temporary network issues and server-side rate-limiting policies, preventing crashes during long-running transcriptions.

Flask Server Watchdog and Force-Kill Utility

• Implementation: The Flask server in api_backend.py now creates a server.pid file on startup. A background watchdog thread monitors for this file's existence. The new standalone kill_server.py script simply deletes the PID file. The watchdog detects the deletion and triggers an immediate, forceful shutdown of the server process.
• Technical Impact: This provides a reliable, cross-platform mechanism to terminate a hung or unresponsive web server, solving issues where the standard Ctrl+C or graceful shutdown fails.

3. Deprecations and Breaking Changes

• Flag Deprecation: The --stream_target_language flag is now deprecated and will be removed in a future version. Users should migrate to --stream_transcribe <language>.
• File Paths: The set_up_env.py script now installs all external tools into a downloaded_assets/ sub-directory. Scripts relying on these tools being in the root directory may need to be updated to use the new ffmpeg_path.bat environment.
• New Dependency for Vocal Isolation: Using --isolate_vocals requires a one-time setup via python set_up_env.py --using_vocal_isolation, which installs Minicoda and associated packages.

---

Your Feedback is Crucial

As this is a beta build, your feedback is invaluable. If you encounter a bug, a crash, or something that doesn't work as expected, please help me improve the software by reporting it.

How to Report an Issue:

1. Go to the Synthalingua GitHub Issues Page.
2. Click on "New Issue
3. Please provide as much detail as possible, including the command you ran and the full console output.

Thank you for being part of this journey and for helping me build a better Synthalingua.