Electronic International Standard Serial Number (EISSN)
Subtitles are a key element to make any media content accessible for people who suffer from hearing impairment and for elderly people, but also useful when watching TV in a noisy environment or learning new languages. Most of the time, subtitles are generated manually in advance, building a verbatim and synchronised transcription of the audio. However, in TV live broadcasts, captions are created in real time by a re-speaker with the help of a voice recognition software, which inevitability leads to delays and lack of synchronisation. In this paper, we present Deep-Sync, a tool for the alignment of subtitles with the audio-visual content. The architecture integrates a deep language representation model and a real-time voice recognition software to build a semantic-aware alignment tool that successfully aligns most of the subtitles even when there is no direct correspondence between the re-speaker and the audio content. In order to avoid any kind of censorship, Deep-Sync can be deployed directly on users' TVs causing a small delay to perform the alignment, but avoiding to delay the signal at the broadcaster station. Deep-Sync was compared with other subtitles alignment tool, showing that our proposal is able to improve the synchronisation in all tested cases.
tv broadcasting; synchronisation; language model; deep neural networks; machine learning