Fact · Voice and AI
79 to 90percent of failures
The τ-Voice authors report that 79 to 90 percent of failures stem from agent behavior rather than from the speech input.
Checked against its source on October 17, 2026
In the source's own words
“qualitative analysis confirms 79--90% of failures stem from agent behavior”
τ-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains, Soham Ray et al., arXiv, 2026, Abstract.
Open the original- What it measures
- The authors analyzed failures qualitatively, after running the benchmark tasks.
- What it does not show
- The authors say this suggests the failures reflect agent behavior under their setup. It is not a measurement of speech recognition alone.