Abstract: ‍ ‍As a common mental health problem， depression seriously affects people’s daily life and even life safety. The detection of depression and depressive mood is meaningful. The common modes of depression detection include EEG， image， text and speech， among which speech signal has the advantages of easy acquisition and less restrictions on use. Therefore， speech-based depression detection research has become a current research hotspot. This paper reviews the latest progress in the field of speech-based depression detection in recent years. Firstly， the depression speech data sets commonly used in current research are introduced， and the methods to deal with the problem of data imbalance are summarized and analyzed. Then， the prosodic features， voice quality features， spectrum-based features and other speech features commonly used in depression speech recognition are summarized， and the characteristics of the features are analyzed. On the other hand， aiming at the problem of small amount of data encountered in depression detection research， the current mainstream few-shot learning methods are briefly described from four aspects： data enhancement， metric learning， meta-learning and transfer learning. Considering the privacy of depressive speech data， the research of depressive speech detection based on federated learning is also introduced， and the data security and edge device deployment are described in detail. Finally， the research status and difficult problems of speech-based depression detection are summarized and prospected.