Public, open-source voice-AI benchmarking for STT and TTS providers. Runs on a schedule against a pinned dataset, writes results to Postgres, and surfaces them at benchmarks.coval.ai. Licensed Apache-2.0 — clone, build, and run with your own provider API keys.

See runner/ for setup and local-run instructions.

This is a research benchmark. Results reflect a specific pinned dataset, a specific set of provider model versions, and a specific normalization pipeline. Absolute numbers may not generalize to production workloads. Latency metrics and per-provider drift over time are the most informative signals. See docs/methodology.md for the full reproducibility contract.

The repository is licensed Apache-2.0; the full warranty disclaimer is in LICENSE §7. Third-party SDK licenses (e.g. provider client libraries) are independent of this repository's license — see each SDK's own license for terms.