Full-Duplex-Bench-v3: First Benchmark for Voice Agent Tool Use Under Real Human Disfluency

arXiv·medium signal

Lin, Chen et al. introduce FDB-v3, a benchmark for evaluating spoken language models under naturalistic speech conditions including five disfluency types, combined with multi-step tool use. Unlike prior work using synthetic speech, this dataset consists entirely of real human audio. Fills a critical gap — voice agents must handle stammering, self-corrections, and interruptions while maintaining tool-use accuracy in production deployments.