Benchmark
AA Speech to Speech Index
Compare native voice models across reasoning, customer-service tasks, human preference and successful tool use.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
A composite index from Artificial Analysis. Version 2.0 began in August 2026; its exact release day is unspecified.
Methodology
Four equally weighted components combine Big Bench Audio, τ-Voice, Speech Agent Arena preference and arena task success. All four results are required. The preference component uses a frozen rating transformed against a fixed baseline. The index combines existing evaluations; it is not an additional task set. Version 2.0 replaced conversational dynamics with arena task success. Frozen index preference differs from live arena Elo.