Understanding features of self-repairs is critical to the exploration of interpreters’ underlying self-monitoring mechanism. As most previous taxonomies just made slight amendments on Levelt’s tripartite taxonomy model — which is designed for monolingual speech production, they failed to match features of interpreting output. In order to resolve the defects of previous classifications, the present paper established an interpreting-tailored taxonomy of self-repairs, which categorized repairs into five major types with altogether nine sub-types. Rather than directly referring to Levelt’s model, each category in the new taxonomy is derived from the descriptive study of data collected from ten interpreting trainees, including their interpreting recordings, retrospection and notes. The qualitative analysis on distribution of each repair category also shows that trainees’ adoption of repairs is mainly triggered by competence deficiency rather than the attempt to facilitate communication.