ASPIRin: Action Space Projection for Interactivity-Optimized Reinforcement Learning in Full-Duplex Speech Language Models

Researchers propose ASPIRin, a reinforcement learning framework that improves full-duplex speech language models by separating turn-taking decisions from semantic generation. The method reduces repetitive output by over 50% compared to standard approaches while maintaining natural conversational dynamics.

ASPIRin addresses a fundamental challenge in conversational AI: enabling machines to engage in natural back-and-forth dialogue without sacrificing output quality. Traditional reinforcement learning approaches that optimize speech timing alongside token generation create a compound optimization problem, leading to degradation in semantic coherence and repetitive patterns. By decoupling when to speak from what to say through Action Space Projection, ASPIRin treats the timing problem as a binary classification task while preserving the language generation pipeline.

This technical innovation emerges from broader efforts to make speech language models more practical for real-world interaction. Full-duplex systems—where both participants can speak simultaneously—require sophisticated turn-taking mechanisms that human conversations naturally employ through subtle timing cues and backchanneling. Previous work struggled to balance responsiveness with semantic quality, forcing researchers to choose between natural interaction patterns and coherent output.

The 50% reduction in duplicate n-grams represents a material improvement in output quality, which directly impacts user experience in conversational applications. Developers building voice assistants, customer service bots, or interactive AI systems benefit from more natural interactions without compromising response accuracy. The framework's effectiveness across multiple turn-taking scenarios suggests broader applicability across different conversational contexts.

The research indicates future speech language models may increasingly rely on modular optimization strategies that isolate distinct aspects of conversation. This approach suggests the field is moving toward more sophisticated architectural designs rather than end-to-end optimization, potentially enabling faster iteration on specific conversational behaviors without full model retraining.

• →ASPIRin decouples timing decisions from language generation, preventing semantic degradation in conversational AI
• →The method reduces duplicate n-grams by over 50% compared to standard reinforcement learning approaches
• →Action Space Projection maps vocabulary to binary states, simplifying the optimization problem for turn-taking
• →Framework demonstrates improvements across turn-taking, backchanneling, and pause handling in full-duplex systems
• →Modular optimization approach suggests future development of more sophisticated conversational AI architectures