# Liberating LLM Capabilities in Full-Duplex Speech Models

Source: [Liberating LLM Capabilities in Full-Duplex Speech Models](<https://www.sota2.com/research/paper/6a279e0890966ade919bf642>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-12T06:00:02.917Z

Authors: Luoyuan Zhang, Bokai Xu, Junbo Cui, Weiyue Sun, Yingjing Xu, Hanyu Liu, Yuan Yao

Published: 2026-05-04T17:54:41.000Z

## 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. Existing work improves 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. We propose Listen-Write-Speak (LWS), a text-first tri-channel paradigm in which 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 under a shared causal attention context. This behavior is implemented entirely through a Token Schema, requiring no architectural modifications, and learned via a two-stage data pipeline that synthesizes per-second cognitive annotations consistent with the revealed input timeline. Empirically, LWS demonstrates strong full-duplex interaction on Full-Duplex-Bench, reaches 4.72 on VoiceBench AlpacaEval, achieves 92.6% writing-speaking consistency, and consistently outperforms its internal ablations on URO-Bench. These results suggest that visible writing can serve as a first-class output channel for speech interaction without sacrificing realtime responsiveness. The code and dataset are available on the project page: https://royalzhang.com/project/lws-page/.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Spoken Understanding and Reasoning on URO-Bench Chinese (Pro)](<https://www.sota2.com/research/sota/spoken-understanding-and-reasoning-on-uro-bench-chinese-pro.md>): Spoken understanding and reasoning; URO-Bench Chinese Pro; U Score: 93.9.

- [Spoken understanding and reasoning on URO-Bench English Pro](<https://www.sota2.com/research/sota/spoken-understanding-and-reasoning-on-uro-bench-english-pro.md>): Spoken understanding and reasoning; URO-Bench English Pro; U Score: 75.4.

- [Duplex Dialogue Turn-Taking on Full-Duplex-Bench](<https://www.sota2.com/research/sota/duplex-dialogue-turn-taking-on-full-duplex-bench.md>): Duplex Dialogue Turn-Taking; Full-Duplex-Bench; Synthetic TOR for Pause Handling: 0.01.

- [Speech-to-Text Reply Quality on VoiceBench AlpacaEval](<https://www.sota2.com/research/sota/speech-to-text-reply-quality-on-voicebench-alpacaeval.md>): Speech-to-Text Reply Quality; VoiceBench AlpacaEval; Quality Score: 4.72.

- [Spoken Understanding and Reasoning on URO-Bench Chinese Basic](<https://www.sota2.com/research/sota/spoken-understanding-and-reasoning-on-uro-bench-chinese-basic.md>): Spoken understanding and reasoning; URO-Bench Chinese Basic; U Score: 79.3.

- [Spoken understanding and reasoning on URO-Bench English Basic](<https://www.sota2.com/research/sota/spoken-understanding-and-reasoning-on-uro-bench-english-basic.md>): Spoken understanding and reasoning; URO-Bench English Basic; U Score: 83.3.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2606.07547>)

[PDF](<https://arxiv.org/pdf/2606.07547.pdf>)

[Implementation](<https://github.com/zly-idleness/lws_demo>)