Automatic video dubbing requires precisely aligned speech, transcripts, and video, but existing datasets often suffer from inconsistent audio quality, unreliable transcriptions, and poor visibility of speakers' faces. These issues make it difficult to train and evaluate dubbing models.

To address this, the paper introduces VoxDub, a bilingual audio-visual dataset designed specifically for lip-synchronized video dubbing. The dataset provides temporally aligned speech, transcripts, and video, aiming to fill a gap in current resources.

The significance lies in offering a more standardized resource for AVD research, potentially improving the reproducibility and comparability of dubbing systems. However, the abstract does not specify which languages are included or the scale of the dataset.