Prime Video uses machine learning models trained to watch video like a human reviewer and automatically flag audiovisual problems — such as freezing, stuttering, or audio that is out of sync — before or during streaming. This work began around 2019 and replaced a largely manual review process.

The Video Quality Analysis (VQA) group trains machine learning models to detect more than 18 types of defects, including frame freezes, pixelation, audio dropouts, and lip-sync failures. A key technical challenge is that defective examples are rare in training data, so the team synthetically injects faults into clean video. One specialized model, called "LipSync," is based on an architecture developed at Oxford University to detect audio-video synchronization errors. In 2023, the team moved its infrastructure from a distributed microservices setup to a consolidated system on Amazon ECS, which cut operating costs by 90% and expanded monitoring from roughly 5% of streams to all customer streams.