Voice agents deployed in structured workflows—such as customer service, healthcare scheduling, and account management—must handle frequent user interruptions while maintaining progress through multi-step procedures. According to a new paper by Salimi, Ahmad, Ma, Wentao, Tang, Yuzhi, Shen, Dongming, Li, Mu, and Smola, Alex, titled "IHBench: Evaluating Post-Interruption Recovery in Voice Agents with Structured Workflows" (arXiv 2606.19595), existing benchmarks for speech-capable models focus only on the timing of interruptions: barge-in detection, endpointing, and turn-taking dynamics. They leave unmeasured what happens after the interruption: does the agent resume the workflow at the correct step? Does it address the user's interjection? Does it avoid re-delivering content the user already heard?
IHBench: A Structured Evaluation Framework
The researchers introduce IHBench (Interruption Handling Benchmark), which evaluates post-interruption recovery in voice agents executing state-machine-driven workflows across 10 enterprise domains. The benchmark injects six interruption types at controlled points mid-utterance, with per-interruption evaluation rubrics generated alongside the data. Each interruption is scored on two axes: task fulfillment and recovery quality.
Key Findings: Closed-Weight Models Lead
The study evaluated 27 audio-language model configurations from OpenAI, Google, and the open-weight community. The results show that models vary widely, and recovery quality depends strongly on the interruption type. Across experiments, closed-weight models are consistently more robust to interruptions than open-weight ones. Specifically:
- Closed-weight models win far more often on task fulfillment.
- They degrade roughly 3.3x more slowly as conversations grow longer.
- They show no audio-versus-text modality gap, whereas open-weight models lose ground on all three metrics.
A table summarising the key comparisons:
| Metric | Closed-Weight Models | Open-Weight Models |
|---|---|---|
| Task fulfillment wins | Far more often | Fewer wins |
| Degradation rate (long conversations) | 3.3x slower degradation | Faster degradation |
| Audio vs. text modality gap | No gap | Gap present |
Validation and Distinct Capability
A human study validated the LLM judge against human annotators, confirming the benchmark's reliability. A cross-benchmark analysis against AudioMultiChallenge indicates that recovery quality is a largely distinct capability axis, not fully captured by existing benchmarks.
Implications for Enterprise Adoption
For CTOs and digital transformation leaders evaluating voice agents for structured workflow domains—such as logistics customer service, healthcare scheduling, or account management—these findings highlight the importance of testing post-interruption recovery. The benchmark provides a standardised way to compare models from major providers like OpenAI and Google against open-weight alternatives. Closed-weight models currently offer superior robustness, especially in long interactions, which is critical for enterprise deployments where interruptions are common and continuity matters. The research also underscores that recovery quality is a separate axis from basic interruption detection, warranting dedicated evaluation in procurement decisions.