AI Subtitle Performance Improvement Notice
Notice
This document is a machine-translated draft and is currently undergoing review. Some content may be inaccurate or differ from the original Korean version. For the most precise information, refer to the Korean documentation.
We sincerely thank all our customers for using the Kollus AI Subtitle service.
Based on your valuable feedback, we have significantly improved the AI Subtitle generation performance to provide more natural and accurate subtitles. Please find the key updates below.
🔧 Key Improvements
Context-aware word selection
- Improved feature: Enhanced performance in selecting the most contextually appropriate word during speech recognition.
- Details: We introduced N-gram, a statistical language model for Korean, to precisely analyze the flow before and after each utterance. Among candidate words that are easily confused due to similar pronunciation, the system now autonomously determines and selects the word that is most natural in the overall context.
Optimized subtitle sync and sentence segmentation timing
- Improved feature: The sync between actual speech and the appearance/disappearance of subtitles is now more accurate.
- Details: We refined the process of merging results from two speech recognition engines. This allows subtitle display timing and sentence breaks (phrase segmentation) to be smoothly adjusted according to the actual speaking speed and flow.
Automatic correction of frequently misrecognized expressions
- Improved feature: The AI automatically corrects specific words and expressions that were repeatedly misrecognized.
- Details: We applied automatic correction rules built by analyzing real-world misrecognition case data. To prevent distortion of otherwise correct sentences, the AI comprehensively cross-verifies the pronunciation and context of a word, and executes automatic correction only when it is safe to do so.
Removal of abnormal subtitles in background noise/silent segments
- Improved feature: Prevents arbitrary subtitles from being generated in segments without voice.
- Details: We resolved the AI hallucination phenomenon in which random text was generated during background noise segments (music or noise only) or silent segments. We strengthened the repeated word detection and removal algorithm so that subtitles are generated accurately only for actual speech segments.