REMIT: Transforming Conversations into Structured Narratives

Working across more than 120,000 hours of video from RTS, this project made at eM+, EPFL, explores how spoken archives can be turned into audiovisual arguments without losing the texture of the original material. Interviews, broadcasts, and testimonies are treated not as content to flatten into a single summary, but as a field of fragments, voices, and recurring motifs that can be brought into relation through montage.

Starting from a curatorial question, the system assembles polyphonic narratives that keep provenance in view. Quotes remain tied to speakers, timestamps, and contexts; distant materials can be placed side by side; and multiple interpretive frames can be tested through the same corpus. Developed for exhibition making, research, and pedagogy, the work turns archival exploration into a visible and discussable process, opening conversations rather than closing them.

Technically, the project combines large language models, agents, retrieval-augmented generation, semantic embeddings, and graph-based indexing to transform unstructured conversations into structured narrative material. The pipeline segments transcripts into distinct topics, enriches each passage with timestamps, entities, metadata, and vector representations, then uses hybrid retrieval, query expansion, and cross-encoder reranking to identify the most relevant excerpts for a given question. A knowledge-graph layer maps relationships between topics, people, places, events, and concepts, allowing the system to generate source-linked collages, subtitles, media trims, and AI-assisted voice sequences that remain accountable to the archive while supporting multilingual interpretation and promptable narrative modes.

‍