Automatic Subtitle Generation for Sound

The emergence of diverse video content over the past decade has highlighted the need for accessibility, particularly for the deaf community who may struggle to understand spoken content without subtitles. This thesis aims to explore and propose an automated method for generating subtitles from audio in videos, thus eliminating the need for intensive user involvement. The methodology is divided into three primary stages: first, separating and converting audio from the video; second, recognizing the speech within the audio; and finally, generating a subtitle file based on the recognized speech. Various implementation strategies for each module are discussed. Experimental results indicate that the current approach yields suboptimal satisfaction levels, suggesting that further refinements are necessary. Proposed improvements include enhancing the parallelization of decoding processes, utilizing better-trained recognition models, and incorporating punctuation into the generated subtitles. These adjustments are crucial for future advancements in automated subtitle generation, ultimately contributing to making video content more accessible to all viewers.