Liberating LLM Capabilities in Full-Duplex Speech Models

A team of researchers has proposed Listen-Write-Speak (LWS), a new paradigm for full-duplex speech models that elevates visible text to a first-class output channel alongside spoken responses, according to a paper posted to arXiv on 4 May 2026 [1]. The LWS framework departs from conventional speech-based large language models, which typically constrain outputs to verbal replies and suppress text-native capabilities such as code generation, structured analysis, and multi-step reasoning during realtime interaction [1][2]. Under the LWS approach, a single autoregressive LLM continuously listens to user audio, writes visible free-form text as its primary output, and speaks a realtime oral response in parallel, all within a shared causal attention context [1][2]. The behavior is implemented entirely through a Token Schema, requiring no architectural modifications to the underlying model [1][2]. The paradigm is learned via a two-stage data pipeline that synthesizes per-second cognitive annotations consistent with the revealed input timeline [1][2]. On the Full-Duplex-Bench benchmark, LWS demonstrated strong full-duplex interaction performance [1][2]. The system reached a score of 4.72 on VoiceBench AlpacaEval and achieved 92.6% writing-speaking consistency [1][2]. It also consistently outperformed its internal ablations on URO-Bench [1][2]. The authors note that these results suggest visible writing can serve as a first-class output channel for speech interaction without sacrificing realtime responsiveness [1][2]. The paper's abstract emphasizes that existing work in the field has improved spoken reasoning or full-duplex turn-taking but still treats text as a hidden intermediate state or a subordinate modality rather than a first-class output channel [1][2]. The LWS model's tri-channel design — listen, write, speak — directly addresses that limitation by making written output inspectable and persistent during live conversations [1][2]. The code and dataset for the project have been made available on the project page [1][2]. The research was submitted under the Computation and Language category on arXiv [1].