Sony launches Woosh AI for professional sound effect generation

Sony's move shifts generative audio from consumer novelty to professional utility. By providing video-conditioned generation and targeting specific foley needs like footsteps or engine sounds, Sony addresses the creative bottlenecks in gaming and post-production where manual sound syncing is costly. This launch materializes Sony's 'Creative Entertainment Vision,' which prioritizes B2B AI tools that assist rather than replace specialized labor. For the broader ecosystem, it signals a bifurcation: generalists like ElevenLabs and Stability AI will compete for creators, while legacy masters like Sony build high-walled, licensed moats for commercial studios. Watch for the public model's integration into digital audio workstations (DAWs) to gauge creator adoption versus existing asset libraries.

Sony AI positioned Woosh as a purpose-built alternative to general-purpose audio models that have largely ignored the sound effects category. The company released the public model with open weights and inference code on GitHub, making it available for non-commercial research, while the private model trained on licensed libraries from Pro Sound Effects and BOOM is available for commercial licensing. Sony AI described Woosh as filling a gap most generative audio tools have overlooked, noting that public audio datasets capture ambient, overlapping recordings while professional libraries contain isolated, purpose-recorded sounds with precise technical annotations. The team is developing a digital audio workstation plugin with planned support for variation generation, inpainting, and personalization features that reflect the granular creative control professionals have requested.

The commercial strategy behind Woosh reflects Sony's broader approach to AI in creative industries. Sony AI released the public model as non-commercial research while pointing the private model toward studio licensing, a bifurcated release model that lets the research community build on the architecture while Sony retains a commercial advantage through proprietary training data. The company has not yet deployed Woosh in production but is working with game studios to evaluate the model and gather feedback, according to the research team. The licensing approach mirrors how Sony has handled other AI tools for creative professionals, keeping high-quality outputs behind commercial agreements.

On technical benchmarks, Woosh's video-to-audio model outperforms comparable baselines on FoleyBench, a dataset designed specifically for visually grounded audio generation, while using fewer parameters. The research paper evaluates Woosh against open alternatives like StableAudio-Open and TangoFlux, showing competitive or better performance across each module on both public and private evaluation data. The distilled variants (Woosh-DFlow and Woosh-DVFlow) enable low-resource operation and fast inference, which matters for integration into real-time production pipelines where latency constraints limit model size.