Full-Duplex-Bench-v3: Benchmarking Tool Use for Full-Duplex Voice Agents Under Real-World Disfluency
This work addresses the need for robust benchmarking of voice agents in naturalistic, disfluent speech scenarios, though it is incremental as it builds on prior benchmarks by focusing on real human audio and multi-step tasks.
The paper tackles the problem of evaluating spoken language models under real-world disfluencies and multi-step tool use by introducing the Full-Duplex-Bench-v3 benchmark, with results showing GPT-Realtime leading in accuracy (Pass@1 of 0.600) and interruption avoidance (13.5%), while Gemini Live 3.1 achieves the fastest latency (4.25 s).
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.