More than Words: In-the-Wild Visually-Driven Prosody for Text-to-Speech
This addresses the dubbing problem for video content creators by enabling automated, synchronized speech generation, though it builds incrementally on existing TTS methods with added visual input.
The paper tackles the problem of generating speech synchronized with video input by introducing VDTTS, a visually-driven text-to-speech model that uses video frames alongside text to produce prosodic variations and timing aligned with the video, achieving synchronization quality approaching ground-truth on benchmarks like VoxCeleb2.
In this paper we present VDTTS, a Visually-Driven Text-to-Speech model. Motivated by dubbing, VDTTS takes advantage of video frames as an additional input alongside text, and generates speech that matches the video signal. We demonstrate how this allows VDTTS to, unlike plain TTS models, generate speech that not only has prosodic variations like natural pauses and pitch, but is also synchronized to the input video. Experimentally, we show our model produces well-synchronized outputs, approaching the video-speech synchronization quality of the ground-truth, on several challenging benchmarks including "in-the-wild" content from VoxCeleb2. Supplementary demo videos demonstrating video-speech synchronization, robustness to speaker ID swapping, and prosody, presented at the project page.