Reverb Open-Source ASR With Diarization for Long-Form Audio

What Happened

An open-source ASR and speaker diarization system named Reverb was posted to Show HN, targeting long-form audio transcription with speaker attribution. The project claims best-in-class open performance for long-form audio, positioning itself against both proprietary APIs and earlier open pipelines that degrade across multi-hour inputs. The release is distributed as open source, with the Show HN thread serving as the initial venue for claims, benchmarks, and community scrutiny.

Why It Matters

Long-form diarized transcription sits at the center of meeting intelligence, media production, podcast indexing, call analytics, and compliance archiving — workflows where audio routinely runs 45 to 180 minutes with overlapping speakers. Until now, teams assembling this capability from open components typically stacked a VAD, a segmentation model, an ASR engine, and a separate diarization model, then absorbed the integration cost, latency, and error compounding at each boundary. Proprietary APIs (AssemblyAI, Deepgram, Rev, Google, AWS Transcribe) abstract this, but at per-minute pricing that scales linearly with volume and creates vendor lock-in on a core data asset. A unified open system that holds quality across long inputs changes the build-versus-buy calculus for any operator processing meaningful audio volume.

Technical Details

Public claims center on long-form performance, which is the failure mode most open stacks handle poorly: speaker drift, timestamp misalignment, and hallucination accumulation over extended context. Diarization quality is typically measured by DER (Diarization Error Rate), with published SOTA on benchmarks like AMI and CALLHOME clustering in the 10–20% range depending on overlap and channel conditions; Reverb's position relative to that band should be verified against its released numbers and evaluation set. Integration profile matters operationally — whether it ships as a Python package, container, or inference server, and whether it requires GPU class hardware for real-time factor targets. Known open-system limitations typically include degraded performance on heavy overlap, code-switching, far-field capture, and domain-specific vocabulary without custom vocabulary injection or fine-tuning hooks.

Operational Impact

The immediate effect is a credible self-hosted baseline for teams currently paying per-minute diarization fees or maintaining brittle multi-model pipelines. For a team processing 10,000 hours monthly, replacing an API at ~$0.10–$0.30/min with self-hosted inference shifts cost from variable to fixed — dominated by GPU amortization — and removes per-seat or per-minute throttling from product design. It also unlocks on-premise and air-gapped deployment for regulated verticals (healthcare, legal, finance, government) that cannot send audio to third-party endpoints. Downstream, transcript pipelines gain a swappable component: teams can A/B Reverb against incumbents on their own audio rather than relying on vendor benchmarks, and rerun historical audio through improved models without renegotiating contracts.