Subglacial seismicity presents the opportunity to monitor inaccessibleglacial beds at the epicentral location and time. Glaciers can beunderlain by rock or till, a first order control on bed mechanics.Velocity-weakening, necessary for unstable slip, has been shown for eachbed type, but is much stronger and evolves over more than an order ofmagnitude longer distances for till beds. Utilizing a de-stiffeneddouble direct shear apparatus, we found conditions for instability atfreezing temperatures and high slip rates for both bed types. Duringstick-slip stress-drops, we recorded acoustic emissions withpiezoelectric transducers frozen into the ice. The two populations ofevent waveforms appear visually similar and overlap in their statisticalfeatures. We implemented a suite of supervised machine learningalgorithms to classify the bed type of recorded waveforms and spectra,with prediction accuracy between 65% - 80%. The Random ForestClassifier is interpretable, showing the importance of initialoscillation peaks and higher frequency energy. Till beds have generallyhigher friction and resulting stress-drops, with more impulsive firstarrivals and more high frequency content compared to rock emissions, butrock beds can produce many till-like events. Seismic signatures couldenhance interpretation of bed conditions and mechanics from subglacialseismicity.