Serverless Text-to-Speech Pipeline

The scenario

A training scenario for a fictional client, Pixel Learning Co., that wanted course content stored in GitHub converted to audio for visually impaired users and for people who learn by listening. The brief was a lightweight, cost-effective pipeline that stays fully serverless.

How it works

• A Python script using boto3 sends speech.txt to Amazon Polly and writes example.mp3.
• Two GitHub Actions workflows run it. on_pr.yml runs when a pull request targets main and uploads to a beta/ folder in S3. on_merge.yml runs on a push to main and uploads to prod/.
• Each run overwrites the file in its folder, and I verified the result in the S3 console, with the AWS CLI and in the Actions log.

What I learned

• Polly’s synchronous synthesis is limited to 3,000 characters. Longer content needs the asynchronous start_speech_synthesis_task method.
• Most of my errors were in the YAML: spelling, indentation and naming consistency across two near-identical workflow files. One of them had the same command typed twice and a missing file name in the production upload.