Voluntary phantom movements are systematically associated with muscle contractions in the residual limb. These latter are specific to the type of movement and can be classified by pattern recognition algorithms. However, phantom mobility generates fatigue that could impact classification metrics. This study explored whether daily phantom movement training at home with no other feedback than inherent somatosensory information can impact the classification success rate. Kinematics and muscle activity were compared between before and after a two-month home training in six major upper limb amputees. Surface EMG patterns were classified to quantify a potential change in the features space with training. Our results showed that this type of training induces faster, smoother, and richer phantom mobility. However, classification metrics did not change with training. When including the new types of movements achievable after training, accuracy did not decrease, indicating that muscle activation patterns associated with these movements were sufficiently different not to interfere with the already existing movement classes. Thus, although phantom training with only somatosensory feedback increases the overall phantom movement capacity, it does not increase the classification success rate. Yet, it is possible that paired with other forms of feedback, phantom training could improve this success rate.