Session transcription that already knows your characters’ names

The fantasy-noun problem, and what actually fixes it

Speech recognition is a prediction problem. A model guesses the most probable word given the sound and everything before it, and "Thordak Ironforge" is not a probable word sequence in any general-purpose training corpus. You get "Thor Dak iron forge", or "thought Dak", and you get it differently on different sessions, so the same NPC ends up under four spellings and your notes stop being searchable.

The fix is not a better general model. It is telling the model what nouns to expect. DM Scribe builds a vocabulary prompt for every transcription from your campaign’s own data: the glossary you curate, your player characters and their aliases, the 200 most recently updated wiki entries, and every nickname recorded in the alias table.

That prompt is capped at 850 characters, which is a real constraint rather than a marketing number — it is the budget the transcription API allows. Recently updated entries win the space, on the assumption that what you touched last is what is currently on screen. If building the prompt fails for any reason, it falls back to empty rather than sending something malformed, and transcription proceeds normally.

What you get back

One transcript per uploaded recording, stored against the session, with per-segment timestamps preserved from the transcription engine. Turn timestamps on to see when something was said; copy the whole thing to the clipboard, or download it as a text file you own.

• Timestamped segments, offset correctly across a split recording
• Readable in the app without downloading anything
• Copy to clipboard, or download as a plain .txt file
• Feeds the wiki, the recap, the timeline and Ask Your Campaign automatically

Long sessions are the normal case, not the edge case

A four-hour session is a large audio file and most transcription APIs will not take it in one request. DM Scribe splits on two levels so you never have to think about it. Sessions over four hours are divided into chunks that are transcribed separately and reassembled under one session; separately, any audio over roughly twenty minutes is split again to fit the transcription API’s per-request limit.

Both are invisible from your side. You upload one file and get one transcript, with timestamps offset correctly across the seams. Transcription runs as a background job in the cloud, so you can close the tab and walk away — you get an email when the transcript, the wiki entries and the recap are all ready.

Audio arrives in whatever your phone, recorder or virtual tabletop produced. It is converted to a normalised WAV before transcription, so you are not converting formats yourself.

If you already have the text, skip the audio entirely

Transcripts from Discord bots, another transcription tool, or a play-by-post game can be imported directly as .txt or .md, or pasted in. The full chain still runs — entity extraction, relationships, recap — and it consumes no transcription hours at all. Re-importing the same text into the same session returns the original transcript rather than running the AI chain twice.

This is a Guild Master feature. It is genuinely the cheapest way to use the product if you already have a transcription pipeline you like and only want the campaign structure built on top of it.

Which plans include it

Included on every plan, including the free Adventurer tier, which gets 2.5 hours a month plus 3 one-time bonus hours at signup. Transcript import requires Guild Master.

Full pricing