Benchmark
Speech Agent Arena
Compare human preferences and successful task completion in live calls with anonymous voice models.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Artificial Analysis runs 35 scenarios: 15 with tools and 20 without them. Participants try the same scenario with two unidentified models.
Methodology
Pairwise choices produce preference ratings. A separate task-success metric checks final tool calls after screening conversation eligibility. Preference and task success measure different outcomes. Participant deviations and unverifiable calls are excluded from task success. This live speech-to-speech arena differs from the text-to-speech voice arenas.