The detection of incipient single-tooth gear faults in gearboxes operating under various speed and load conditions is investigated. The problem is of high importance as the timely detection of critical gear faults may enhance safety and significantly reduce maintenance costs. The effects of the considered faults on the observed gearbox dynamics are minor and largely masked by those effects of the various rotating speeds and loads, thus leading to a challenging fault detection problem. The study focuses on exploring the potential of two methods, both based on Multiple Model framework: A non-parametric Order Spectrum (OS) based, and a parametric AutoRegressive (AR) model-based. Both involve angular resampling of random vibration signals via computed order tracking using the tachometer signal from a reference gearbox shaft. Their performance is assessed with the gearbox in the healthy state, as well as under two distinct levels of incipient single-tooth pinion fault using thousands of experiments run under 21 rotating speeds and 4 different loads. The results demonstrate the methods' effectiveness, as well as their superiority over an alternative approach based on Wavelet Packet Decomposition.