Reverb ASR Model: Open-Source Long-Form Audio with Diarization

Reverb, an open-source ASR model with integrated diarization, was released via Show HN with claims of state-of-the-art long-form audio performance. It combines transcription and speaker labeling in a single pass, targeting media, meeting, and call-analysis workloads.

For operators running transcription pipelines, this consolidates what typically requires two separate models—ASR and diarization—into one dependency. That reduces integration complexity and latency, particularly for asynchronous batch jobs processing hours-long recordings. The open-source license also removes per-minute API costs for high-volume or privacy-sensitive data, shifting the cost center to self-hosted compute.

The immediate operational change: teams can now evaluate a single model for end-to-end speaker-attributed transcripts, potentially retiring separate diarization services and their alignment glue code. Second-order effect: as long-form accuracy improves in open weights, proprietary transcription APIs face pricing pressure on exactly the workloads where they historically justified premiums—meetings, legal depositions, and support calls. Builders should benchmark Reverb against their current stack on real, noisy audio before assuming parity, but the barrier to switching has just lowered.