Myoelectric controllers for upper limb prostheses are susceptible to signal disturbances across practical conditions. In particular, electrode liftoff or wire breakage introduce interface noise that, even if only present in a single channel, is detrimental to controller performance. We trained a supervised denoising variational autoencoder to learn a low-dimensional subspace underlying muscle activation patterns that was robust to noise in single EMG channels. Two latent space classifiers, which used the deep learning model, and two conventional LDA-based classifiers were used to classify wrist and hand gestures from clean and synthetically corrupted EMG signals. The baseline LDA classifier, trained on clean data only, suffered a marked increase in errors when evaluated on the corrupted data. The second LDA classifier, trained on clean and corrupted data, improved robustness to noise. Regardless, both latent space methods significantly outperformed both LDA methods in classifying clean and corrupted data. These results highlight that interface noise has adverse effects on current pattern recognition controllers but that deep learning inspired latent space classifiers can mitigate these effects and achieve highly accurate movement classification.