Modern cellular networks utilising the long-term evolution (LTE) and the coming 5G set of standards face an ever-increasing demand for low-latency mobile data from connected devices. Header compression is employed to minimise the overhead for IP-based cellular network traffic, thereby decreasing the overall bandwidth usage and, subsequently, transmission delays. We employ machine learning approaches for the prediction of Robust Header Compression version 2's (RFC 5225) compression utility for VoIP transmissions, which enables the compression to dynamically adapt to varying channel conditions. We evaluate the prediction models employing R^2 and mean square error scores next to complexity (number of coefficients) based on an RTP specific training data set and a separately captured live VoIP audio call. We find that the proposed weighted Ridge regression model explains about 70% of the training data and 72% of a separate VoIP transmission's utility. This approach outperforms the Ridge and first-order Bayesian regressions by up to 50% and the second and third order regressions utilising polynomial basis functions by up to 20%, making it well-suited for utility estimation.