Dominik’s Radio Times brings together recordings of the programmes Dominik Lukeš presented and produced for Future Radio and Reading4U: interviews, talk, music and the reasons he gave on air for choosing each record.
How it started
The site began as a test of how open models, run locally, can be used to process a multimodal archive: hours of speech and music on old recordings, turned into something people can search, read and listen to.
What the models did
- Transcription. The recordings were mainly transcribed with Qwen3-ASR 1.7B, an open speech-recognition model; five supplementary transcripts were made with Parakeet TDT 0.6B v3. A separate audio classifier marked the stretches of music, so song lyrics were not transcribed.
- Summaries. Gemma 4 (the 31B model, 30.7 billion parameters) read the transcripts and wrote the summaries and topic lists on the show and feature pages.
All these models ran locally on Dominik’s own hardware, an NVIDIA DGX Spark. No audio was sent to an online service.
The main transcription batch took about 8 hours. The later, more detailed Gemma pass over 508 documents and 1,136 transcript chunks took about 17 hours 19 minutes, running from 24 to 25 October 2026. Earlier passes were separate and are not counted here.
The processing was supervised by Codex, powered by GPT-6.1 Sol, which was also used to confirm some broadcast dates from email records.
What to expect
Speech recognition makes mistakes, especially with names, and the timings of passages may be slightly out. The summaries are only as accurate as the transcripts they were made from.
What you can hear
The site plays selected clips and music samples extracted from the recordings, not whole broadcasts.