Tusk's Tomes

Import a transcript

The simplest route in: you already have the text, and Tomes turns it into a chronicle. Nothing to install, no GPU, no Python.

Tomes doesn't care where the text came from — a previous Whisper run, a manual notetaker, a NotebookLM export, a Discord chat log, a Roll20 or Foundry log, or a transcript someone else produced.

The short version

  1. Open the Chronicle tab.
  2. Paste your transcript into the Raw transcript box.
  3. Set a campaign name and session number, then Begin the Chronicle.

That's the whole route. Everything after it is the same six-phase pipeline the other two routes feed.

What good input looks like

The pipeline works from whatever it's given, but it works better with some structure:

  • Speaker labels help enormously. A line like [Kaziel] I open the door is far easier to attribute than a bare stream of prose. If your source has them in any form, keep them.
  • Timestamps are optional. They give the chunker better segmentation boundaries, but plain text is perfectly workable.
  • Length is not a problem. The chunker sizes input to whichever model each phase is routed to; a three-hour session is ordinary.
Streamlined

A glossary matters more than the transcript's polish. Phase 1 grounds names against your glossary and lore, so a transcript riddled with misheard fantasy names still comes out right — provided the correct spellings are in the Tome of Lore. Fix a name once and it stays fixed for every future session.

On an AMD or Intel GPU? Read this before you go hunting

If you're here because the built-in transcriber won't use your card, this is the relevant background — it only accelerates on NVIDIA, and it's natural to assume some other Whisper build will use your card. Here's what I found when I actually went looking, to save you the same afternoon:

ProjectPrebuilt GPU backendsVerdict
whisper.cppCPU, BLAS (still CPU), CUDAIts Vulkan backend exists in source but is not in any release build. You'd have to compile it yourself with the Vulkan SDK.
sherpa-onnxCPU, CUDANo DirectML build shipped.
Const-me/WhisperDirectCompute — any DirectX 11 GPUGenuinely works on AMD and Intel. Windows-only, and no release since July 2023.

So the only ready-made thing that will actually use an AMD or Intel card is Const-me/Whisper, and it's been unmaintained for years. If you're willing to accept that, it produces plain text you can paste straight in via this route.

Streamlined

Compiling whisper.cpp with GGML_VULKAN=1 yourself works well, and Tomes has a whisper.cpp bridge built in for exactly that case. It detects your build, reads its capability line so it can tell you outright if you've ended up with a CPU-only binary, and falls back to the built-in engine rather than failing a run. Nothing is downloaded; you supply the build and the model.

If neither appeals, importing YouTube captions needs no GPU at all and is the route I'd point most people to.

Next