# Tracing summaries to a radio archive

- Researcher: Dominik Lukeš
- Role: Led the AudioArchive project and presented the radio programmes it catalogues
- Field: digital archiving
- Tools: Codex (GPT-6.1 Sol); Qwen3-ASR 1.7B; Gemma4 31B
- Published: 2026-10-24
- Source: https://audioarchive.semanticmachines.fyi/
- Written with the current flagship models (Claude Opus 5.5 and GPT-6.1 Sol) available.

## What I did

I used Codex and local models to build a searchable catalogue of my radio recordings. The catalogue links summaries to passages in the transcripts so the wording and attribution can be checked.

I set the questions for the radio archive, supplied knowledge of the collection, and judged which results were useful enough to extend. GPT-6.1 Sol in Codex coordinated the work, built processing tools, and helped design and run checks. Qwen3-ASR transcribed the recordings locally, and Gemma4 31B worked on the transcripts locally to produce summaries and catalogue entries. Cloud agents reviewed selected text. The original recordings stayed on local machines.

We compared Gemma and Qwen on the same ten interviews and five show recordings. Each model produced three drafts out of fifteen that passed the source-matching checks. We then had the model select transcript segment IDs. Code copied the exact transcript words and attached their timestamps. Records that failed the checks were held back and the original outputs were kept.

## What I learned

Matching the excerpt to the transcript tells me where the words came from. I still have to check whether the transcript captured the words spoken and whether the summary attributes them to the right person. A sample review found that one summary had assigned my first-person memories to my guest. We saved editorial corrections separately from the model output.

## Limits of this account

This is an account of my own archive project. The pilot counts concern source matching in fifteen drafts per model. I have no independent replication to report here. This entry draws on the sections of the public case study covering the team, model roles, and checks on summary evidence. The case study is dated 24 October 2026 and was updated on 25 October. GPT-6.1 Sol wrote most of it, with edits and additional information from me.
