How to transcribe a D&D session

Length is the first wall. Otter's free plan stops at 30 minutes per conversation and allows three file imports in the lifetime of the account. Its Pro plan stretches to 90 minutes per meeting and only the Business plan reaches four hours. A normal session is longer than what most of these products consider a long meeting.

Crosstalk is the second. In a meeting people take turns. At a table three players react to a natural 1 at once, and a transcriber has to decide who said what. YouTube's own help page lists overlapping speakers as a reason automatic captions fail, alongside videos that are too long, which is a fair description of every session ever recorded.

Names are the third and the worst. Speech recognition is trained on real words, and your campaign is full of invented ones. Strahd becomes Strod, Vallaki becomes the lackey, and the wizard's name comes out differently every time it is spoken. The words around the name are fine, and the one word you will later search for is wrong.

Whisper is OpenAI's open-source speech recognition model, released under the MIT licence. It runs on your own computer, so the audio never leaves it, and it costs nothing but electricity. It is a command-line tool that needs Python and ffmpeg, and it writes the transcript as plain text, subtitle files, or JSON.

It comes in six sizes, from tiny to large, plus a faster turbo model that is the default. The size you can run depends on your graphics card: the README lists roughly 1 GB of video memory for the smallest models, about 6 GB for turbo, and about 10 GB for large. It will run without a graphics card, slowly, and four hours of audio is a long time to wait on a laptop processor.

What Whisper does not do on its own is tell speakers apart. Its documentation describes no speaker labels, so you get one unbroken stream of text. WhisperX, a maintained open-source project built on top of it, adds word-level timestamps and speaker labels. The price is more setup: a CUDA toolkit if you want the graphics card to do the work, and a free Hugging Face account and access token for the speaker model.

This is the right route if you are comfortable in a terminal and privacy matters more than convenience. Speakers come back as numbered labels rather than names, and nothing in the toolchain knows a player from a character.

Otter is the one most people try first. The free Basic plan gives 300 minutes a month with the 30 minute cap, which rules out a session. Pro is $16.99 a user per month, or $8.33 on the annual plan, with 1,200 minutes, 90 minutes per meeting, and ten file imports a month. Business is $30 a month, or $19.99 annually, and is the first plan that takes a four hour recording in one piece.

Descript is an audio and video editor that transcribes as a side effect, which makes it a good fit for a table that publishes an actual play. The free plan includes one media hour a month. Hobbyist is $24 a month, or $16 billed annually, for 10 media hours, and Creator is $35, or $24 annually, for 30. It can transcribe a multitrack recording with each speaker on their own track, and exports the transcript as .docx, .txt, .rtf, .md, or .html.

Rev sells human transcription at $1.99 a minute. For a four hour session that is $477.60, which is the right price for a published show and the wrong one for a weekly home game. Its machine transcription comes with a subscription: a free plan with 45 minutes a month, then $29.99 a seat per month for 5,000 minutes.

All three return a transcript and stop. None of them has heard of an initiative order.

Speech recognition is also sold directly, by the minute, and the prices are startling next to a subscription. Deepgram lists its Nova-3 model for pre-recorded audio at $0.0043 a minute, which is $1.03 for four hours, with $200 of free credit to start and speaker labels billed as an add-on. AssemblyAI lists $0.21 an hour plus $0.02 an hour for speaker labels, about 92 cents for the same session, with a free tier of up to 185 hours.

What you get back is JSON, not a document. You write the script that uploads the file, turns the response into something readable, and stores it. If that sounds like a pleasant Sunday, this is the cheapest reliable way to transcribe a campaign. We use Deepgram ourselves for the transcription step, as our privacy policy says.

Check what your recorder can do before paying for anything else. Craig, the Discord recording bot, includes transcripts on its $8 a month Patreon tier, delivered as WebVTT, SRT, or plain text. If the recording is already there, that is one less upload.

Phones transcribe too. Voice Memos shows a transcript on an iPhone 12 or later in a short list of languages, and Google's Recorder on a Pixel transcribes on the device, though its speaker labels are US English only. Both are built for a memo, not a four hour table, so treat the result as a search aid.

Uploading the recording to YouTube as a private video and downloading the automatic captions is the classic free trick. It works sometimes. YouTube's help page warns that captions may never appear for a long video or for overlapping speakers, and tells you to review whatever you get.

Suppose every tool above worked perfectly. You now own 30,000 accurate words a week, and the question you had on Thursday, what did we promise the dragon, is somewhere in the middle of them. A transcript is the raw material. The thing you wanted was the recap, the list of names, and the two decisions that mattered.

RollSummary is built around that second step. You upload the audio or video and get the transcript with speaker labels and timestamps, and it proposes which speaker is which member of your party. The names already in your campaign, including any spelling you have corrected before, are passed to the recognition step, so a name you fixed once tends to stay fixed. It detects the language your table speaks from the audio, across more than 30 languages, and writes everything back in it.

Then it writes the notes: a structured recap with a TL;DR, the key events in order, and the open threads, plus a wiki note for each person, place, and quest, linked to each other. Nothing is saved until you approve it. The transcript stays searchable afterwards, and the GM's campaign chat can quote the exact moment something was said, with its timestamp. A four hour session takes about ten minutes.