Note Start here: structured, source-linked capability map across public voice-agent benchmarks.

Note Speech-based tool tasks, multi-turn interaction, and safety.

Note Listening, speaking, viewing, multi-turn behavior, and safety.

Note Multilingual customer-service ASR across five locales.

Note Multi-turn spoken-agent suite; public files were limited at the 2026-09-01 review.

Note VoiceAgentBench paper.

Note VoiceAssistant-Eval paper.

Note tau-Voice full-duplex benchmark paper.

Note Spoken instruction-following, knowledge, reasoning, safety, accent, and limited multi-turn evaluation.

Note Audio2Tool benchmark paper for evaluating spoken tool use and conversational repair.

Note VoiceBench paper covering spoken instruction following, safety, reasoning, and robustness.

Note Spoken tool-call benchmark with eight tiers spanning multi-intent, correction, multi-turn, and overlapping speech.

Note EVA-Bench paper on end-to-end task accuracy, speech fidelity, conversation quality, and robustness.

Note 213 end-to-end enterprise voice-agent scenarios with accuracy, experience, perturbation, and reliability evaluation.