Candidate Speech Extraction from Multi-speaker Single-Channel Audio Interviews

Video interviews are increasingly common, initially due to travel restrictions during the pandemic and now extended for economic reasons. However, in many developing nations, limited bandwidth makes video interviews impractical, leading to a preference for telephonic interviews. These interviews are often recorded for audit and analysis purposes. A candidate's performance in an interview depends not only on their knowledge but also on how they respond to the interviewer. Both the content of their responses and their communication skills influence the overall selection process. For any downstream interview analysis or automation, a reliable method is needed to identify speech segments spoken by the candidate in a multi-speaker, single channel interview conversation. In this paper, we propose a pipeline to accurately identify candidate speech segments. This pre-processing step is crucial for analyzing various aspects of a candidate's interview performance, such as answer analysis, confidence level, and emotional expression.