Google’s New Video-to-Audio Generator Watches Pixels to Create Soundtracks

Deep Mind V2A AI research

Deep Mind showed off the latest results from its generative AI video-to-audio research on Tuesday. It’s a novel system that combines what it sees on-screen with the user’s written prompt to create synced audio soundscapes for a given video clip.

The V2A AI can be paired with vide -generation models like Veo, Deep Mind’s generative audio team wrote in a blog post, and can create soundtracks, sound effects, and even dialogue for the on-screen action. What’s more, Deep Mind claims that its new system can generate “an unlimited number of soundtracks for any video input” by tuning the model with positive and negative prompts that encourage or discourage the use of a particular sound, respectively.

The system works by first encoding and compressing the video input, which the diffusion model then leverages to iteratively refine the desired audio effects from background noise based on the user’s optional text prompt and from the visual input. This audio output is finally decoded and exported as a waveform that can then be recombined with the video input.

The best part is that the user doesn’t have to go in and manually (read: tediously) sync the audio and video tracks, as the V2A system does it automatically. “By training on video, audio and the additional annotations, our technology learns to associate specific audio events with various visual scenes, while responding to the information provided in the annotations or transcripts,” the Deep Mind team wrote.

The system is not yet perfected, however. For one, the output audio quality is dependent on the fidelity of the video input and the system gets tripped up when video artifacts or other distortions are present in the input. According to the Deep Mind team, syncing dialogue to the audio track remains an ongoing challenge.

“V2A attempts to generate speech from the input transcripts and synchronize it with characters’ lip movements,” the team explained. “But the paired vide- generation model may not be conditioned on transcripts. This creates a mismatch, often resulting in uncanny lip-syncing, as the video model doesn’t generate mouth movements that match the transcript.”

The system still needs to undergo “rigorous safety assessments and testing” before the team will consider releasing it to the public. Every video and soundtrack generated by this system will be affixed with Deep Mind’s SynthID watermarks. This system is far from the only audio-generating AI currently on the market. Stability AI dropped a similar product just last week while ElevenLabs released their sound effects tool last month.