SearchMinutes made the public record of civic meetings searchable by transcribing the meetings themselves. Local government business is largely conducted in recorded public meetings, and the recordings are hours of video with no transcript or index.
The project was shut down in 2026 and is no longer active.
My part
SearchMinutes was a Michigan Radio (NPR) project. I came in as a consultant late in its life, as it was being prepared for transfer to the Associated Press. The main job was rebuilding the transcription pipeline: the existing design sent recordings to a transcription service, and I replaced that with OpenAI's Whisper running inside the pipeline. Alongside that were the application changes needed for the handover to AP.
How it works
The Minutes pipeline ran on Google Cloud Platform, driven by a suite of configurations for monitoring civic media feeds. Newly discovered media were injected into a pipeline of queued processing stages that registered the media with the system, then downloaded and transcribed it. Transcripts were written to a search index used by the SearchMinutes website.
The registry and queues were handled by Redis. The system was written in Python, with Whisper providing the transcripts.