Abstract: ‍ ‍Depression is a common mental disorder， and early detection and diagnosis is essential for depression prevention and treatment. Speech depression detection is an efficient and convenient tool in the current computer-aided detection methods. In order to explore whether different emotional stimuli have an impact on speech depression detection， this paper firstly constructed an acoustic feature set for depression analysis， followed by analyzing the significant differences in acoustic features between depression and non-depression under different emotional stimuli using a non-parametric test， and then used an experimental design with a combination of emotional stimuli （positive， negative， neutral） and task types （text reading， question-and-answer） to build speech depression detection models through machine learning and deep learning algorithms. The experimental results demonstrate that the use of emotional stimuli affects the depression detection task to some extent and that negative emotional stimuli are more likely to trigger depression-related emotions and to have an impact on the individual’s pronunciation characteristics， thus achieving better detection performance than positive stimuli and neutral speech.