Abstract: Continuous autoregressive speech generation directly models speech representations in continuous space, and has attracted increasing attention. Nevertheless, challenges remain in modeling probability distributions in continuous space and balancing generation quality with inference efficiency. In recent years, extensive research has been devoted to building high-quality and efficient continuous autoregressive speech generation models. This paper provides a detailed classification and in-depth analysis of existing methods from three perspectives: continuous representation methods, probability distribution modeling approaches, and model architecture design. Trends in this field are discussed and the relevant literature is systematically reviewed.