How AudioShake Uses AWS to Separate Any Sound at Scale

AudioShake is an AI company tackling one of sound’s most persistent technical challenges: once audio is mixed, separating its components back into individual tracks has historically been impossible at any useful quality level. The San Francisco-based startup serves the music industry, film studios, sports broadcasters, and AI training companies—anywhere that valuable sound data is locked inside complex, layered recordings that existing tools cannot untangle.

The scale of the problem is substantial. A song recorded in 1965 exists as a single blended track—vocals, drums, guitar, and bass fused together. A live sports broadcast layers commentary, crowd noise, and copyrighted music in real time. Film studios attempting to dub content for international markets often cannot remove the original dialogue cleanly. AI companies training multimodal models on real-world speech struggle when audio inputs are noisy and overlapping. Standard signal processing rules cannot handle this complexity, leaving the problem effectively unsolved until deep learning changed the calculus.

AudioShake built proprietary separation models trained on large licensed audio datasets using Amazon EC2 G6 GPU instances for high-performance training and inference. Amazon S3 handles the massive file volumes that audio processing demands, while Amazon ECS and EKS provide the container orchestration layer needed to scale workloads across different use cases. AWS Lambda and Step Functions coordinate the multi-step processing pipelines that convert raw audio uploads into separated stems. AWS CloudFormation manages infrastructure as code for rapid deployment and iteration. The result is three production separation models: one for music stems, one for dialogue, music, and effects separation, and one for multi-speaker voice isolation.

The impact of this infrastructure has been demonstrated through partnerships that would have been impossible with traditional audio tools. AudioShake helped Green Day recover guitar tracks from a 1991 recording for a TikTok remix, enabled the remastering of classic live jazz albums from the estates of Oscar Peterson and Nina Simone, and separated the complete audio of The Wizard of Oz so the film could be re-released in Dolby Atmos at The Sphere in Las Vegas. The company also won the top prize at the 2024 AWS re:Invent Unicorn Tank competition.

AudioShake describes its broader objective as building the “intelligence layer of sound”—the equivalent transition that made text searchable and programmable. With AWS as its infrastructure backbone, the company is extending from individual track separation into use cases including ALS patient voice restoration, call center audio clarity, sports broadcast copyright management, and AI training data generation for the next generation of multimodal models.