Voice benchmarks

Voice benchmarks.Measured on speech that matters.

Our proprietary voice benchmarks test voice models on held-out speech, so results reflect how a model will behave in use, not how well it fits public test sets.

Voice benchmarks · reportSL
Held-out voice suite
Never published
ModelNoiseAccentsTurn-takingLatency
Model A
Model B
Model C
Model D
Expert review

Model B misreads drug names in noisy calls. 14 examples attached.

LNLinguist · reviewer
Illustrative interface

Why proprietary

Public test sets leak into training data over time, and scores on them stop meaning much. Our benchmark data is held out and never published, so a model cannot have seen it.

How an evaluation runs

You give us access to your model. We run it on the benchmark, review the results with experts, and send a report with scores and the examples behind them. Your model and results stay confidential.

Questions, answered.

Do you publish leaderboards?

Not today. Results are shared with the model owner only.

Every safe model
has an expert behind it.