v0.4.4 Dramatically Improved Alignment Success Rate

🎉 Overview

LattifAI Python v0.4.4 is a performance-focused release that brings significant improvements to alignment success rate and speaker name handling.

---

✨ Key Highlights

🚀 Dramatically Improved Alignment Success Rate

This release includes a major enhancement to the beam search decoding parameters, resulting in significantly higher alignment success rates, especially for challenging audio scenarios.

What Changed:

• search_beam: Increased from 50 to 200 (4x improvement)
• output_beam: Increased from 20 to 80 (4x improvement)

Impact:

• ✅ Higher Success Rate: Dramatically reduced alignment failures for complex audio
• ✅ Better Path Exploration: Wider beam search explores more decoding paths
• ✅ Improved Accuracy: More robust alignment for difficult audio segments
• ✅ Enhanced Reliability: Better handling of noisy or low-quality audio

🔧 Fixed Speaker Name Parsing & Restoration

Resolved critical issues with speaker name detection and text restoration in subtitles.

What Was Fixed:

• Regex Pattern Fix: Corrected SPEAKER_PATTERN2 to properly match uppercase speaker names
• Format Support: Enhanced support for speaker formats like JOHN DOE: text or SPEAKER: text
• Text Restoration: Fixed speaker name restoration when exporting aligned subtitles

Impact:

• ✅ Accurate Speaker Detection: Properly identifies speaker names in various formats
• ✅ Preserved Speaker Info: Maintains speaker attribution through alignment pipeline
• ✅ Better Format Support: Works with diverse subtitle speaker naming conventions