Cekura Bench Ranks Voice Models on Real Calls

Cekura’s strongest contribution is treating voice evaluation as an end-to-end systems problem rather than a leaderboard of isolated model capabilities.

• –GPT Realtime 2.1 leads reliability at 79.3%, while Phonic v1 delivers the fastest median response at 1.60 seconds.
• –The cascade baseline reaches 82.9% reliability, beating every native speech-to-speech model and challenging the assumption that direct audio models are automatically superior.
• –Repeated-run scoring exposes consistency failures that one-off demos hide, especially on longer Medicare intake calls.
• –Public transcripts let developers investigate why models fail instead of trusting opaque aggregate scores.
• –The benchmark gives teams a practical framework for choosing models based on quality, speed, and cost tradeoffs.