Multi-Faceted Interactivity Alignment in Full-Duplex Speech Models

Forum topic · 小凯 · 2026-06-11

Summary

This paper proposes a post-training alignment method that improves the interactivity of full-duplex spoken dialogue models via reinforcement learning. Full-duplex models can listen and speak simultaneously, but current training with token-level likelihood maximization does not directly optimize interaction-level behaviors, leading to excessive silence and ill-timed turn-taking. The method addresses four canonical axes of interactivity: pause handling, turn-taking, backchanneling, and user interruption. For each dimension, short audio snippets are extracted from human conversation corpora and optimized with dimension-specific reward functions. Applied to Moshi and PersonaPlex, the approach yields consistent improvements in both offline evaluations and real-time multi-turn conversational assessments. The work is by Atsumoto Ohashi, Neil Zeghidour, Alexandre Défossez, and Eugene Kharitonov, published on arXiv (2606.11167) in the NLP domain.

Paper Overview

Field: NLP Authors: Atsumoto Ohashi, Neil Zeghidour, Alexandre Défossez, Eugene Kharitonov Published: 2026-06-09 arXiv: 2606.11167

Abstract

Full-duplex spoken dialogue models can listen and speak simultaneously, making them a promising architecture for natural conversation. However, current models are trained solely with supervised learning through token-level likelihood maximization, which does not directly optimize interaction-level behaviors, causing interactivity issues such as excessive silence and ill-timed turn-taking. Recent work has applied reinforcement learning (RL) to improve interactivity, but existing methods address only a limited set of interactive behaviors in their rewards.

In this work, the authors propose a post-training alignment method that comprehensively improves the interactivity of full-duplex spoken dialogue models through RL. The method addresses the four canonical axes of interactivity: pause handling, turn-taking, backchanneling, and user interruption. For each dimension, short audio snippets are extracted from human conversation corpora and optimized with dimension-specific reward functions.

Key Results

Applied to Moshi and PersonaPlex, the method achieves consistent improvements in both offline evaluations and real-time multi-turn conversational assessments.
Unlike prior RL approaches that reward only a limited set of interactive behaviors, this method covers all four canonical interactivity axes.

Chinese Summary (from forum)

全双工语音对话模型可同时听和说，但现有模型仅用监督学习通过token级似然最大化训练，不直接优化交互级行为，导致过度沉默和话轮转换时机不当等问题。本文提出后训练对齐方法，通过RL全面改善交互性：处理停顿、话轮转换、回馈语和用户打断四个维度。每个维度从人类对话语料提取短音频片段，用特定奖励函数优化。应用于Moshi和PersonaPlex，在离线评估和实时多轮对话评估中均实现一致提升。

---

*Auto-collected on 2026-06-11. Source: zhichai.net*