Full-Duplex-Bench-v3: Benchmarking Tool Use for Full-Duplex Voice Agents Under Real-World Disfluency

2026-04-06

AI Summary

This work addresses the limitations of current speech agents in handling disfluent utterances in real-world spoken interactions and the lack of systematic evaluation of their multi-step tool-use capabilities. The authors propose the first full-duplex speech agent benchmark based on authentic human speech, annotating five types of spoken disfluencies and introducing cross-domain chained API tasks to comprehensively assess model performance across accuracy, latency, and turn-taking control. Six representative architectures—including GPT-Realtime, Gemini Live, Ultravox, and a cascaded Whisper–GPT-4o–TTS pipeline—are evaluated. Results show that GPT-Realtime achieves the highest first-pass success rate (0.600) and lowest interruption rate (13.5%); Gemini Live 3.1 exhibits the shortest latency (4.25 s) but only a 78.0% turn acceptance rate; and the cascaded system demonstrates perfect turn control at the cost of high latency (10.12 s).

Abstract

We introduce Full-Duplex-Bench-v3 (FDB-v3), a benchmark for evaluating spoken language models under naturalistic speech conditions and multi-step tool use. Unlike prior work, our dataset consists entirely of real human audio annotated for five disfluency categories, paired with scenarios requiring chained API calls across four task domains. We evaluate six model configurations -- GPT-Realtime, Gemini Live 2.5, Gemini Live 3.1, Grok, Ultravox v0.7, and a traditional Cascaded pipeline (Whisper$\rightarrow$GPT-4o$\rightarrow$TTS) -- across accuracy, latency, and turn-taking dimensions. GPT-Realtime leads on Pass@1 (0.600) and interruption avoidance (13.5\%); Gemini Live 3.1 achieves the fastest latency (4.25~s) but the lowest turn-take rate (78.0\%); and the Cascaded baseline, despite a perfect turn-take rate, incurs the highest latency (10.12~s). Across all systems, self-correction handling and multi-step reasoning under hard scenarios remain the most consistent failure modes.

Problem

• full-duplex
• voice agents
• disfluency
• tool use
• spoken language models

Innovation

• full-duplex voice agents
• disfluency handling
• tool use benchmarking
• real-world speech
• multi-step API chaining