The work of John Antonakis and colleagues at UNIL has identified markers in the transcripts of speeches that are indicative of charisma. Charisma in turn has been shown to be an important high level semantic, able to predict the number of views a YouTube video will receive, or even how well a US presidential candidate will perform. The Deep Charisma project aims to apply deep learning to the automatic prediction of charisma. We use techniques from document classification to automatically assign charismatic markers to texts. These in turn correlate quite highly with hand annotations of the same texts. The finished system will allow such classification on a much wider variety of texts and speeches, at much lower cost, and without the variability of human transcription.