AudioToText: Turning Meeting Recordings into a Tool for Requirements Management

The project AudioToText was born from the practice of implementing BI projects, where analysts conduct interviews with client stakeholders and collect requirements. A common problem is that different people may express contradictory requirements for the same entity, and without a structured approach, these contradictions are often missed, leading to inconsistent technical specifications. AudioToText aims to turn meeting recordings into a structured artifact that can be reviewed, compared, and used for requirement alignment. The project uses MLX Whisper for speech recognition, speaker diarization, NLP processing, and a Django web interface for reviewing and exporting results. It generates summaries in the meeting × speaker × entity perspective, and supports manual NER annotation for quality evaluation. The project is designed to run locally on Apple Silicon to keep sensitive data on-premises. Future plans include using PostgreSQL and pgvector for semantic comparison of summaries to automatically detect conflicts in requirements. The project also distinguishes between users, project members, and speakers, with a role-based access model for viewers, analysts, and project managers.