Preprint
τ-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains
Soham Ray et al., arXiv, 2026
Read the originalRead on October 17, 2026
- What it is
- A benchmark that gives voice agents multi-turn tasks with policies and tools, runs them with simulated callers who interrupt and use accents, and compares the results with text performance on the same tasks.
- How much weight it carries
- Preprint, submitted March 14, 2026, not yet peer reviewed. The simulated users and tasks are the authors design, so scores describe this benchmark rather than every real call.
Facts taken from this source
- The τ-Voice authors report that 79 to 90 percent of failures stem from agent behavior rather than from the speech input.
“qualitative analysis confirms 79--90% of failures stem from agent behavior” (Abstract)
- On the τ-Voice benchmark, GPT-5 with reasoning completed 85 percent of tasks, against 31 to 51 percent for voice agents on clean audio.
“while GPT-5 (reasoning) achieves 85%, voice agents reach only 31--51% under clean conditions” (Abstract)
- On the τ-Voice benchmark, voice agents completed 26 to 38 percent of tasks under realistic conditions, with noise, diverse accents and interruptions.
“and 26--38% under realistic conditions with noise and diverse accents” (Abstract)