Reverb ASR+Diarization: Best open-source ASR for long-form audio

What Happened

Reverb, an open-source automatic speech recognition (ASR) system with integrated speaker diarization, was posted to Hacker News as a Show HN, reaching 12 points. The project targets long-form audio transcription, combining speech-to-text and speaker separation in a single pipeline rather than requiring separate chained services. It is positioned as a free alternative to commercial ASR APIs such as those from major cloud providers.

Why It Matters

Commercial ASR pricing scales linearly with audio volume, which imposes a structural cost penalty on any application that processes long meetings, interviews, or continuous multi-speaker audio. Reverb removes this per-minute billing model entirely, converting a variable operating expense into a fixed compute cost that builders can amortize. The integrated diarization matters more than the transcription itself for meeting analysis: most production pipelines today chain a cloud ASR endpoint with a separate speaker-identification service, doubling latency and failure surface. For voice agent builders, this consolidation reduces the number of external dependencies in the critical path and eliminates a category of vendor lock-in. On-premise and air-gapped deployment becomes practical, which addresses privacy constraints in legal, healthcare, and government workflows where audio cannot leave the network.

Technical Details

Reverb couples ASR with speaker diarization in a unified inference path, avoiding the serialization overhead of running two models against the same audio stream. The architecture is optimized for long-form input, meaning it handles extended context windows and speaker-turn boundaries that short-utterance models typically mishandle. Open-source availability means weights and inference code can be run locally on GPU or CPU depending on throughput requirements; exact benchmark numbers (word error rate, real-time factor, diarization error rate) are not specified in the initial post and should be validated against your own audio distribution before committing. Limitations to verify: language coverage, handling of overlapping speech, robustness to far-field or noisy capture, and whether streaming inference is supported or only batch. Integration typically requires a serving layer (e.g., a FastAPI or gRPC wrapper) plus enough VRAM to hold the model for the target concurrent-stream count.

Operational Impact

For teams running meeting transcription or voice agents, the day-to-day shift is from managing API quotas, retry logic, and per-minute cost dashboards to managing GPU utilization and model versioning. Cost modeling changes shape: instead of a linear bill per hour of audio, you provision fixed compute that handles N concurrent streams, which makes high-volume or always-on transcription economically viable where it previously was not. Pipeline architecture simplifies — one service replaces two — reducing the number of places where audio must be serialized, transmitted, and stored. Privacy-sensitive deployments that previously required redaction or contractual controls before sending audio off-prem can now keep raw audio inside the perimeter. The tradeoff is operational burden: you now own uptime, scaling, and model updates that a cloud vendor previously absorbed.