spoken-language-models

本文识别了全双工语音语言模型中的"状态惯性"，即在用户打断时，模型的内部预测焦点滞后，并提出了一种无需训练的激活引导方法来改善打断处理。

This paper introduces InterRS, a method for real-time speech generation that interleaves reasoning steps during natural pauses in speech, achieving better performance on math and logic benchmarks while maintaining fluent and instant responses.