WavFlow: Audio Generation in Waveform Space

WavFlow is a new audio generation framework that works directly in waveform space, rather than relying on latent compression.

It uses patchification and amplitude lifting with flow matching to synthesize high-fidelity sound.

Trained on 5 million video-text-audio triplets, it was evaluated on VGGSound and AudioCaps.

WavFlow matched or outperformed strong latent-based baselines on video-to-audio and text-to-audio tasks.

The result suggests competitive audio synthesis may be possible without a latent bottleneck, simplifying multimodal model design.