Facial animation consists of breathing life into computer graphic characters. The major challenge of facial animation is to fool the human's eyes, very acute at recognizing natural motion. The emergence and increasing availability of motion capture technologies have opened a new era, where realistic animation generation is more deterministic and repeatable. The theoretical promise of MoCap is the ability to capture and retarget human performances completely. In reality, even professional motion capture setups often fall short of a perfect animation. Animation editing is therefore unavoidable and frequently the bottleneck of modern performance-based animation pipelines, requiring a considerable amount of time and special skills. We propose to improve the facial animation signal editing process by leveraging the recent deep learning techniques. These methods enable producing realistic motion sequences, endowed with all the specificities included in the real animation used for the training. We adopt this technology to enhance the cleaning facial animation editing process, providing a solution preserving and even restoring the facial motion dynamics. We also explore new possibilities to synthesize facial animations from alternative high-level control inputs, more semantic and intuitive than the current standard animation parameters. Pushing forward the motion editing toward user controllability, we provide a robust and interactive regressive method to modify a facial animation efficiently from temporally dense parameters. To bypass the limitations of regression algorithms, we propose a generative framework suitable for unsupervised and supervised motion editing.