Pattern recognition-based myoelectric control schemes aim to classify surface electromyographic (sEMG) signals generated by skeletal muscles for use in prosthetics or wearable robots. Many factors impact the EMG signal quality and reliability, such as motion artifacts, electrode shift, and reduced conductivity over time, which calls for robust pattern recognition-based myoelectric control schemes. The manifold hypothesis – a breakdown of high-dimensional space to a lower-dimensional representation to explain how the higher-dimensional space operates – provides a framework to discover the representation or manifold of multimodal biological signals. This paper presents a pattern recognition-based myoelectric control scheme that creates compressed latent space representations using a contrastive variational autoencoder (VAE) with an integrated classifier. The VAE model was designed, trained with a secondary dataset of hand gestures, and validated subjectwise as well as for the group. The individual subject data yielded distinct and separable latent spaces with high classification accuracy ranging from 75.1\% to 91.7\%. Further, the model was optimized to improve classification accuracy, reaching 87.45\% to 97.49\%. The model architecture was generalizable across the subjects, and the compressed latent space achieved high performance when the representation was separable and distinct. The proposed VAE model and latent space representation demonstrate its feasibility and utility for use in myoelectric controls.