Call holding time (CHT) is a statistical indicator in a cellular network, directly affecting network performance and providing critical insight for the network service provider. CHT distribution estimation literature relies on the classical estimation theory that targets to determine parameters of a function. Hence, related work can be considered as making use of parametric approaches. However, the required assumptions for these approaches may not be correct for obtaining an accurate model. In this paper, we introduce a probabilistic framework for CHT distribution estimation, which makes use of Dirichlet process mixture of lognormal distributions. The purpose of this work is to provide a practical Bayesian inference framework to enable the extraction and identification of user behaviors, which are not available through traditional estimation techniques. The performance of the proposed framework is tested on a large data set that is obtained from a mobile switching center of a cellular network service provider composed of calls from Global System for Mobile Communications (GSM) and High-Speed Packet Access (HSPA) networks. Accuracy of the obtained CHT distributions is verified through several performance tests, showing that all distribution estimates have significance levels of 0.99.