Lexicap at home

In 2022 Andrej Karpathy downloaded all ~322 episodes of the Lex Fridman podcast, ran them through Whisper, and published the transcripts as a site he called Lexicap - raw VTT included for anyone else who wanted them.¹

The point was not the site. The point was the instinct: a podcast locked in audio is not searchable, not quotable, not linkable. Transcribe the whole thing and it becomes text you can actually work with.

Lexicap took a cloud pipeline and a lot of manual work to convert one show. That is fine for a one-off by someone who wanted to prove a point.

It does not scale to the ten shows you actually follow. Nobody hand-runs a back-catalogue through a cloud transcriber every week - the effort, and the per-minute bill, stop you before you start.

Vocateca does the same audio-to-text step on-device. Subscribe to a show by its Apple, Spotify or RSS address and it transcribes new episodes as they publish, and backfills the archive behind them.

A whole back-catalogue is just disk space and a night of compute - no upload, no per-minute fee. Timestamps and speakers survive, and the output exports straight to Markdown, SRT, VTT or JSON.

Karpathy named the underlying problem in the same breath: so much information is locked up in opaque audio files.² A local transcript library answers that - accessible, searchable, and yours.

Audio behind an app can be taken down or moved. A folder of plain-text transcripts you own does not rot, greps in a second, and is the substrate everything downstream - a knowledge network, an LLM query, a citation - is built from.