Various studies suggest that marmosets (Callithrix jacchus) show behavior similar to that of humans in many aspects. Analyzing their calls would not only enable us to better understand these species but would also give insights into the evolution of human languages and vocal tract. This paper describes a technique to discover the patterns in marmoset vocalization in an unsupervised fashion. The proposed unsupervised clustering approach operates in two stages. Initially, voice activity detection (VAD) is applied to remove silences and non-voiced regions from the audio. This is followed by a group-delay based segmentation on the voiced regions to obtain smaller segments. In the second stage, a two-tier clustering is performed on the segments obtained. Individual hidden Markov models (HMMs) are built for each of the segments using a multiple frame size and multiple frame rate. The HMMs are then clustered until each cluster is made up of a large number of segments. Once all the clusters get enough number of segments, one Gaussian mixture model (GMM) is built for each of the clusters. These clusters are then merged using Kullback-Leibler (KL) divergence. The algorithm converges to the total number of distinct sounds in the audio, as evidenced by listening tests.