The paper evaluates the use of large language models (LLMs) as judges for assessing conversational voice agents, comparing human judgments with GPT‑4.1 and GPT‑5 across telecom and retail interactions. It examines agreement, metric‑level correlations, and consistency across three evaluation configurations (p0, p1, p2) to determine how reliably LLMs can judge conversational quality and safety. The study finds that LLM‑based evaluation can be effective but its reliability varies by metric and configuration, suggesting a hybrid approach where LLMs handle scalable assessment while humans focus on metrics requiring contextual interpretation.
By Anupam Purwar, Shashank Singh, Kritika Srivastava
CallScreenBench is a benchmark for evaluating small, on-device language models that act as phone secretaries, focusing on their ability to handle unknown inbound calls without a cooperative task. The benchmark measures owner endorsement through five call-and-note metrics, each paired with counter-metrics and uncertainty estimates, and includes guardedness diagnostics to identify safe, tool‑free proxies. Results across 4‑bit checkpoints of 0.6‑4 B parameter models show varying performance on service, recall, plausibility, and triage discrimination, highlighting trade‑offs between quality and guardedness.
By Jiaqi Gan, Haoyuan Tang, Jamey Z. Liang, Siying Chen, Ankit Raj, Kidus Zewde, Yuchen Zhou, Yuxin Zhang, Simiao Ren
SpeechGym is an audio‑native environment that lets two omni‑modal models converse entirely in native audio, eliminating external ASR/TTS and API boundaries while preserving the tasks, tools, and success checks of a standard text‑based agent benchmark. By training end‑to‑end, the framework addresses perceptual failures—such as misheard arguments that cascade into failed calls—and behavioural failures, both of which are automatically labeled for free. Using per‑turn process rewards to overcome reward sparsity, agents trained in SpeechGym transfer to an independent voice benchmark, doubling task success and improving efficiency in turns and tokens.
By Jiajun Fan, Jingyuan Li, Prashanth Gurunath Shivakumar, Jia-Hong Huang, Qi Luo, M. Maruf, Ivan Bulyko, Ge Liu, Roger Ren
The study evaluates conversational speech models by placing them at 427 pauses and 482 turn ends from 176 unscripted conversations, scoring their responses on timing, latency, length, wording, and recall of recent dialogue. Hosted APIs (GPT Realtime, Gemini 3.1 Flash Live, Grok Voice) respond at almost every pause, with latency ranging from 1–4 s for silence-based endpointers and about 8 s for semantic turn detection, while full‑duplex models (PersonaPlex, Moshi) remain silent more often and can interrupt within human reaction time. Recall rates vary: hosted systems remember 41–60 % of facts under a Claude Opus 5.5 judge and 19–26 % with a strict word matcher; full‑duplex models recall 17–22 % and 12–15 % respectively.
By Arya Bastani, Iman Modarressi
The paper introduces a reproducible framework that converts existing text-based tool‑calling benchmarks into audio‑based evaluations using text‑to‑speech, speaker variation, and environmental noise, while preserving original annotations. Experiments on seven omni‑modal models show that performance varies strongly by model and task, with notable gaps between text and voice results. The study also includes a stress test, a reference‑free LLM‑as‑judge protocol, and demonstrates that open‑source judges can match proprietary ones, aiding privacy‑preserving evaluation.
By Md Tahmid Rahman Laskar, Xue-Yong Fu, Seyyed Saeed Sarfjoo, Quinten McNamara, Jonas Robertson, Shashi Bhushan TN
arXiv:2610.34653v1 Announce Type: new
Abstract: Latency-sensitive spoken agents must know not only which tool to call, but also when they have heard enough to act. Waiting for the turn to complete de...
By Yizhou Peng, Chen Chen, Donghang Wu, Hexin Liu, Eng Siong Chng