Hearing aids are tasked with the undesirable job of compensating an impaired, highly-nonlinear auditory system. Historically, these devices have either employed linear processing or relatively unsophisticated, nonlinear processing techniques. With increasingly more accurate models of the auditory system, expanding computational power, and many more objective measures which utilize these models, we are at a turning point in hearing aid design. Although subjective listener tests are often the most accepted methods for evaluating the quality and intelligibility of speech, they inherently treat the auditory system as a "black box." Conversely, model-based objective measures typically treat the auditory system as a cascade of physical processes. As a result, objective measures have the potential to provide more detailed information about how sound is processed and about where and why quality or intelligibility breaks down. Provided that we can generalize model-based objective measures, we can use the measures as tools for understanding how to best process degraded signals, and therefore, how to best design hearing aids. However, generalizability is a key requirement. Since many of the well-known objective measures have been developed for normal-hearing listeners in the context of audio codecs, we are unsure about the generalizability of these measures to predicting quality and intelligibility for hearing-impaired listeners with "unknown" datasets (i.e. a set on which it was not trained) and distortions which are specific to hearing aids. Relatively recently, however, Kates and Arehart (Journal of the Audio Engineering Society, 2010) proposed the Hearing Aid Speech Quality Index (HASQI), which is a model-based objective measure that predicts quality for normal-hearing and hearing-impaired listeners by taking into account many of the distortions which hearing aids introduce. HASQI solves many of our concerns of generalizability for predicting quality, but it still remains to test HASQI's ability to predict quality with datasets on which it was not trained. Thus, we explore the robustness of HASQI by testing its ability to predict quality for "unknown" de-noised speech, and we directly compare its performance to some other metrics in the literature.