Language agnostic missing subtitle detection

Abstract Subtitles are a crucial component of Digital Entertainment Content (DEC such as movies and TV shows) localization. With ever increasing catalog (≈ 2M titles) localization expansion (30+ languages), automated subtitle quality checks becomes paramount. Being manual creation process, subtitles can have errors missing transcriptions, out-of-sync blocks with the audio incorrect translations. Such erroneous result in an unpleasant viewing experience impact viewership. Moreover, correction is laborious, highly costly requires expertise languages. A typical process consists (1) linear watch movie, (2) identification time stamps associated blocks, (3) correcting procedure. Among three, taken to entire movie by human expert most consuming step. This paper discusses problem transcription, where corresponding some speech segments DEC non-existent. We present solution augment automatically identifying timings non-transcribed dialogues language agnostic manner. The step then be performed either human-in-the-loop mechanism or using neural transcription (speech-to-text same language) translation (text-to-text different languages) engines. Our method uses voice activity detector (VAD) classifier (AC) trained explicitly on corpora for better generalization. three steps: first, we use VAD identify (predicted blocks). Second, refine those AC module removing leading trailing non-speech identified VAD. Finally, compare predicted dialogue file (subtitle blocks) flag transcriptions. empirically demonstrate that proposed (a) reduces 10%, (b) improves 2.5%, (c) false positive rate (FPR) overextending 77%, (d) block-level precision 119% over baseline human-annotated dataset blocks.