Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Guan-Ting Lin, Chen Chen, Zhehuai Chen, Hung-yi Lee

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.

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

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 (WhisperGPT-4oTTS) -- 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.

Paper Structure

This paper contains 30 sections, 1 figure, 8 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Full-Duplex Evaluation
  4. Tool-Use Voice Agents
  5. Benchmark Design
  6. Task Domains and Mock APIs
  7. Dataset Preparation and Difficulty Tiers
  8. Audio Data Collection and Demographics
  9. Experimental Setup
  10. Evaluation Metrics
  11. Results
  12. Overall Performance
  13. Performance by Difficulty
  14. Robustness to Disfluency
  15. Performance by Domain
  16. ...and 15 more sections

Figures (1)

  • Figure 1: The Full-Duplex-Bench-v3 framework for evaluating real-time voice agents. The diagram illustrates a full-duplex interaction where the agent continuously processes user speech containing natural dysfluencies, such as fillers and self-corrections. Full-Duplex-Bench-v3 comprehensively assesses the agent across three primary dimensions: (1) Tool-use Performance (accuracy of API selection and execution), (2) Turn-Taking Dynamic (handling interruptions and conversational flow), and (3) Latency Breakdown (timing from listening to tool execution and speech generation).