Alibaba Tongyi Launches PrismAudio: Audio-Visual Sync, Sound Effects Follow

Alibaba Tongyi Lab launches the video-to-audio framework PrismAudio, focusing on ambient sound and sound effect synthesis. The model pioneers the combination of reinforcement learning and chain-of-thought, using a "decomposed chain-of-thought" to analyze video content, timing, sound quality, and spatial position before generating audio. It introduces four "teachers" (semantic, temporal, aesthetic, spatial) to score and optimize separately.

2 days ago · PrismAudio is a video-to-audio (V2A) framework launched by Alibaba Tongyi Lab, capable of automatically adding ambient sound effects to silent videos. The model pioneers "decomposed chain-of-thought" technology, allowing the model to first consider sound content, timing, texture, and spatial position...

March 24, 2026 · We introduce PrismAudio, the first framework to integrate Reinforcement Learning into V2A generation with specialized Chain-of-Thought (CoT) planning. Our approach decomposes...

Filmmakers & Editors Recording foley takes days. PrismAudio gives me a working draft in seconds that I can refine or use directly.

March 26, 2026 · Tongyi Lab open-sources PrismAudio, a powerful video-to-audio (V2A) model. With only 518M parameters and 0.63s inference latency, it comprehensively surpasses existing 5B-scale methods across five metrics: semantic, temporal, aesthetic, spatial, and subjective scoring. Its core...

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.