Plenty of apps will transcribe a lecture recording. The problem shows up three weeks later: you have eleven recordings, a quiz on Friday, and a question no single transcript can answer — in which lecture did the professor actually explain backpropagation, and what exactly did she say?

Transcription is a feature. A searchable semester is a library. The difference is what happens after the transcript exists.

The local version of this workflow

Drop your recordings into a folder and open it in StashBase, an open-source desktop app:

  1. Download a speech model once (Settings → Transcription). Three sizes: Tiny (74 MiB), Base (141 MiB), or the default Small (465 MiB).
  2. Transcription runs entirely on your machine. No per-minute API fees, no audio leaving your laptop — which matters, because lecture recordings contain other people’s voices and often fall under course recording policies. You can cancel and rerun while viewing a file.
  3. Every recording becomes timestamped text. Video works too — the audio track is extracted and transcribed (StashBase is not a video editor, and doesn’t try to be).
  4. Search the whole semester at once. Keyword search works offline out of the box; results point to the file and the moment, so you land at the right minute of the right lecture instead of scrubbing. Turn on search by meaning and it finds the concept even when you don’t remember the professor’s exact words.

The folder becomes the study system

Because it’s just a folder, the transcripts sit next to everything else from the course — your notes, the slides, the assigned papers — and the whole thing is one searchable space. Connect it over MCP and you can ask Claude directly: “find where lecture 7 covers eigenvalues, then summarize it against my notes.” The agent searches the same index, scoped to the folders you chose, and its answer points back to real timestamps you can verify.

The same recipe works beyond lectures — the standout case is transcribing your own mock interviews for system design prep, where the recordings are yours and the review pass is the whole point.

Honest notes before you build your workflow on this

  • Model size is a real trade-off: Tiny is fast and rough, Small is the default for a reason. Re-running with a bigger model is one click.
  • Transcription quality on heavily accented or noisy recordings is what you’d expect from on-device speech models — check a sample before trusting a whole semester to it.
  • StashBase is Alpha, Apache-2.0, on macOS, Windows, and Linux. Install it, point it at this term’s folder, and find out what week three actually said.