From Text to Voice: A Reproducible and Verifiable Framework for Evaluating Tool Calling LLM Agents

TL;DR

This paper introduces a framework to convert text-based tool-calling benchmarks into audio-based evaluations for voice agents, enabling reliable assessment of speech tool use without re-annotating datasets.

Contribution

It presents a dataset-agnostic, reproducible method using text-to-speech and environmental noise to evaluate multimodal models on audio tool-calling benchmarks.

Findings

Model performance varies significantly across tasks and models.

Degradation mainly due to misunderstandings of argument values in speech.

Open-source judges with 8B+ parameters align well with human preferences.

Abstract

Voice agents increasingly require reliable tool use from speech, whereas prominent tool-calling benchmarks remain text-based. We study whether verified text benchmarks can be converted into controlled audio-based tool calling evaluations without re-annotating the tool schema and gold labels. Our dataset-agnostic framework uses text-to-speech, speaker variation, and environmental noise to create paired text-audio instances while preserving the original dataset annotations. Based on extensive evaluation of 7 omni-modal models on audio-converted versions of Confetti and When2Call, our framework demonstrates that the performance is strongly model- and task-dependent: Gemini-3.1-Flash-Live obtains the highest Confetti score (70.4), whereas GPT-Realtime-1.5 performs best on When2Call (71.9). On Confetti, the text-to-voice gap ranges from 1.8 points for Qwen3-Omni to 4.8 points for GPT-Realtime-1.5. A targeted analysis of failure cases demonstrates that degradations most often reflect misunderstandings of argument values in the speech. Considering real-world deployment scenarios, we further report text-only results, an ambiguity-based reformulation stress test, and a reference-free LLM-as-judge protocol validated against human preferences. Notably, we find that open-source Qwen3 judges with at least 8B parameters exceed 80% agreement with proprietary judges, supporting privacy-preserving evaluation. Overall, our framework provides a verifiable and reproducible first-stage diagnostic that complements purpose-built audio corpora.