ScrAIbe started as an internal tool for converting laboratory interviews and sensor briefings into searchable text. Off-the-shelf speech-to-text services struggled with German-English code switching and overlapping speakers, so I built a pipeline that combines state-of-the-art ASR with probabilistic diarisation and confidence scoring.
Key pieces include:
- A modular inference stack (Whisper + Pyannote) orchestrated through containerised workers so labs can run the service on-premises.
- A calibration layer that flags low-confidence passages and surfaces timestamps for quick review.
- Automated QC reports that let researchers jump straight to the segments that need manual corrections. The project is open source because reproducible infrastructure should not be a black box. You can read the documentation, run the containers locally, or extend the diarisation modules for your own corpora.