← Back to research

Research Brief · 2026-11-02 Supporting

InteractionBench: Measuring Timing, Task Accuracy, Whisper and Volume, and Grounding in Real-Time Voice Systems

The idea, in brief

One harness plays the same recorded audio to seven voice systems and scores when they speak, wait and stay quiet. None corrected grammar in time.

Editorial summary · not the author abstract. Read the linked paper for the complete abstract and claims.
Figure 1: the evaluation harness, from recorded user audio through each system’s driver to shared graders. Cropped from PDF page 2, because the HTML figure is not an image file.
Figure 1: the evaluation harness, from recorded user audio through each system’s driver to shared graders. Cropped from PDF page 2, because the HTML figure is not an image file. Original figure · paper authors ↗

Release evidence

Public harness and graders under Business Source License 1.1; per-scenario audio is in a linked Google Drive folder. Not rerun.

Last checked · 2026-11-06