Related physiological studies have shown that Mel-frequency cepstral coefficient (MFCC) is a discriminative acoustic feature for depression detection. This fact has led to some works using MFCCs to identify individual depression degree. However, they rarely adopt neural network to capture high-level feature associated with depression detection. And the suitable feature pooling parameter for depression detection has not been optimized. For these reasons, we propose a hybrid network and lp-norm pooling combined with least absolute shrinkage and selection operator (LASSO) to improve the accuracy of depression detection. Firstly, the MFCCs of the original speech are divided into many segments. Then, we extract the segment-level feature using the proposed hybrid network, which investigates the depression-related information in the spatial structure, temporal changes and discriminative representation of short-term MFCC segments. Thirdly, lp-norm pooling combined with LASSO is adopted to find the optimal pooling parameter for depression detection to generate the utterance-level feature. Finally, depression level prediction is accomplished using support vector regression (SVR). Experiments are conducted on AVEC2013 and AVEC2014. The results demonstrate that our proposed method achieves better performance than the previous algorithms.