arXiv:2606. 07547v1 Announce Type: cross Abstract: Speech-based large language models are typically constrained to spoken replies, which limits their user-facing outputs to what can be verbalized and suppresses text-native capabilities such as code generation, structured analysis, and multi-step reasoning in realtime interaction, for tasks that require persistent, structured, and inspectable intermediate outputs.
The paper introduces a frontend‑backend architecture for full‑duplex speech‑to‑speech models that enables tool calls while preserving natural conversational flow. The frontend emits a delegation token and streams ASR transcripts to a text‑based backend LLM, which performs tool calls and returns results that are re‑injected into the frontend via a lightweight prefill‑and‑repeat mechanism before streaming TTS synthesis. In single‑turn evaluations the system achieves 92‑97% tool‑call recall, 81.2% accuracy in rejecting irrelevant calls, and competitive performance on Full‑Duplex‑Bench‑V3 and EVA‑Bench when paired with a large backend model.
arXiv:2610.20172v1 Announce Type: cross Abstract: Full-duplex speech systems must decide when to enter, hold, or relinquish the conversational floor, even as multiple speakers overlap or backchannel...
arXiv:2609.21967v1 Announce Type: cross Abstract: We introduce NemotronLabs VoiceChat, an open full-duplex speech-to-speech model with native tool-calling capabilities. NemotronLabs VoiceChat combine...
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.
The paper introduces a pipeline that generates intent‑labeled, two‑channel conversational speech from relational event lists, enabling controlled synthesis of full‑duplex dialogue with 42 phenomena across eight families in English and Mandarin. By having an LLM author each event’s speaker, text, conversational act, and attachment, and then aligning and timing these events independently, the system produces diverse, realistic turn‑taking signals. Experiments show that models trained on this synthetic corpus achieve higher floor‑occupancy accuracy and better start‑speaking/listening F1 scores compared to models trained on prior data.
Full-duplex speech models require training data that preserves turn-taking, overlap, interruption, and backchannel behavior, yet these signals are entangled across speakers in noisy real-world recordi...
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...
The paper introduces llmovoice, a middleware that explicitly models voice context for large language model (LLM) serving in voice AI applications. By incorporating speaking rate, background noise, packet loss, and other paralinguistic factors into a bounded context, llmovoice guides the LLM to generate more aligned responses. Experiments show significant reductions in speaking‑rate errors, false interruptions, and model usage costs, especially in long voice sessions.
The paper investigates how to evaluate audio‑language models by separating the use of acoustic evidence from the need to invoke a generative audio model. Using a controlled call‑decision framework, the authors compare policies that rely on transcript labels, encoder outputs from CLAP, AST, or WavLM, and optional calls to generative models such as Qwen2‑Audio, Qwen2.5‑Omni, or MOSS‑Audio. Results on the VocalSound dataset show that while transcript‑only accuracy is low (0.296), encoder‑only controls achieve high accuracy (≈0.85) without any generative calls, and adding generative calls yields only a marginal improvement (0.925 vs. 0.921).
arXiv:2609.38867v1 Announce Type: new Abstract: Large language model (LLM) computer-use agents are typically evaluated with clean written instructions, despite speech being an increasingly popular in...
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.