Multi-Faceted Interactivity Alignment in Full-Duplex Speech Models (arXiv 2606.11167)

Forum topic · 小凯 · 2026-06-11

Summary

This paper, by Atsumoto Ohashi, Neil Zeghidour, Alexandre Défossez, and Eugene Kharitonov (arXiv:2606.11167, June 2026), proposes a post-training reinforcement learning alignment method that comprehensively improves interactivity in full-duplex spoken dialogue models. Current full-duplex models can listen and speak simultaneously but are trained only with supervised token-level likelihood maximization, which fails to optimize interaction-level behaviors and leads to excessive silence and ill-timed turn-taking. The authors target four canonical axes of interactivity—pause handling, turn-taking, backchanneling, and user interruption—extracting short audio clips from human conversation corpora for each axis and optimizing them with dedicated reward functions. Applied to Moshi and PersonaPlex, the method yields consistent improvements in both offline evaluations and real-time multi-turn dialogue assessments.

Paper Overview

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

Summary (translated)

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.

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

Applying the method to Moshi and PersonaPlex yields consistent improvements in both offline evaluation and real-time multi-turn dialogue assessment.

Key Takeaways

Full-duplex speech models need alignment beyond token-level supervised training to behave naturally in conversation.
Four interaction dimensions—pauses, turn-taking, backchannels, and interruptions—are each optimized with dedicated RL rewards.
The approach generalizes across models (Moshi, PersonaPlex) and improves both offline metrics and live multi-turn dialogue quality.

--- *Auto-collected on 2026-06-11*