Amazon AWS cuts video accessibility cost with generative AI audio description system
Using Amazon Nova, Rekognition, and Polly, AWS engineers demonstrated an automated pipeline that generates narrated audio descriptions for video content—potentially replacing a process estimated at $25 per minute when done manually. Still an early experiment, not a production deployment.
Background
The scale of inaccessible video content online is significant, with the majority of publicly available video lacking audio descriptions. The primary obstacle is cost: the current manual production workflow involves multiple specialists and averages more than $25 per minute, according to the International Documentary Association. AWS engineers set out to explore whether generative AI could automate this process end-to-end.
What Was Implemented
- Amazon Rekognition Segment API used to partition video files into individual scenes by detecting shot boundaries and technical cues (black frames, color bars)
- Amazon Nova Pro model (accessed via Amazon Bedrock) applied to each video segment to generate descriptive scene-by-scene narration text
- Amazon Polly used for text-to-speech conversion, producing a final MP3 audio description file
- Source video and output artifacts stored in Amazon S3; processing run on Amazon EC2 or Amazon SageMaker notebooks
- Workflow described in pseudocode and guidance; not a packaged production product
Results
The post reports no measured outcome KPIs such as cost reduction, time savings, or quality scores. The stated goal is to demonstrate the technical feasibility of end-to-end automation. The authors note that "this end-to-end automation has the potential to significantly reduce the time and cost required to make video content accessible for visually disabled audiences." The $25-per-minute manual baseline is cited from a third-party source (International Documentary Association). The system itself remains in an early experimental stage as of June 2025.
Lessons
- Video segmentation before AI analysis improves transcription accuracy; large models benefit from chunked inputs rather than full-video processing
- Prompt engineering is critical — without explicit instructions, models may produce preamble ("In this video…") that degrades audio quality
- Function calling / structured outputs in Amazon Bedrock can be used to enforce consistent scene-description format
- Iterative human review of scene descriptions before Polly synthesis is recommended to catch errors
- Even an early-stage experiment surfaces a compelling ROI argument: any per-minute cost below $25 represents improvement over the manual baseline