SAGE

Criminal activities are increasingly coordinated via telecommunication channels. Telephone fraud generates extensive audio data that make manual analysis impossible. The SAGE project responds to these growing demands on the Public Prosecutor’s Office of the Canton of Zurich and the Zurich Forensic Science Institute (FOR) for speaker number estimation and speaker diarization (who speaks when).

SAGE develops robust speaker embeddings, i.e., computational voice model, specifically optimized for forensically challenging conditions: children’s and female voices, poor audio quality, and complex multi-speaker conversations. The system enables users to provide feedback to iteratively improve results.