Monitoring runner exertion in real-time can provide unique insights that help improve training and reduce injuries. Most existing methods use heart rate (HR) as a physiological proxy of it, but this does not always correspond to self-perceived exertion. This is an additional factor in determining overall strain and is typically evaluated with the Borg rating of perceived exertion (RPE) scale. In recent years, speech has been one of the many modalities used to monitor exertion; however, mostly used to predict physiological measures using speech collected after a physical task. In this work, we contrast the manifestation of subjective vs objective exertion on speech signals obtained while running in real-life environments. We identify and interpret a set of prosodic and spectral features related to both markers, and proceed to train deep learning models that directly predict RPE and HR from speech, obtaining an average Pearson correlation of .341 and .418, respectively.