How to Transcribe a D&D Session: Live vs. Upload

Every D&D group has had the conversation. Someone asks what the lich's name was, or whether the party actually agreed to that deal with the fey, and four people remember four different things. Notes help, but somebody has to take them, and that person usually wants to be playing instead.

Transcribing your sessions fixes this, and it has become much easier in the last couple of years. It is also easy to do badly: a muddy recording, a transcript full of misheard elf names, and a forty-page wall of text nobody reads. This guide walks through what actually matters, the four realistic ways to get a transcript, and how to make it something your table uses rather than something that sits in a folder.

Full disclosure up front: we make Realms of Shod, which is one of the four options below. We have tried to describe the other three the way we would describe them to a friend, including where they are the better choice.

What do you need before you can transcribe a D&D session?

A usable recording. Every method below, including ours, is limited by the audio you feed it. Transcription models have become very good, but they still cannot recover a line that was mumbled into a pizza box while two other players were talking over it.

For online games, you are in good shape almost by default. Discord, and most voice tools, give every speaker their own audio stream, which makes it far easier to tell players apart. The main things that go wrong are:

Push-to-talk players who forget to push. Voice activity detection is kinder to transcripts.
Background noise. A fan, a mechanical keyboard, or a TV in the next room gets transcribed as mush and can drown out quiet speakers. Discord's noise suppression helps.
Music bots. If you play ambient music through a bot, it is one more "speaker" for the transcriber to sort out. Keep it low.

For in-person games, the recording is the hard part. A single phone in the middle of the table records everyone at once, so the transcriber has to work out who is talking from the sound alone, a job called speaker diarization. It works reasonably well with four or five clear voices and gets worse as more people talk over each other. A few habits make a big difference:

Put the recorder in the middle of the table, not next to the DM, who is already the loudest person there.
Get it off the table surface if you can. A cheap stand or a folded towel cuts the thud of dice rolls and cups being set down.
An inexpensive USB conference microphone beats a phone in most rooms. A dedicated field recorder beats both, but is overkill for most groups.
Do a thirty-second test at the start of the night and listen back once. It is much better to find out the mic is muffled before a four-hour session than after.

One more thing to decide early: do you want the transcript during the session, or after it? Most tools only work after the fact, from a finished recording. A few work live, while people are talking. Live is useful if players want to scroll back mid-session ("what exactly did the riddle say?"), and it means the notes exist the moment you stop. After-the-fact is simpler and works with any recording, including sessions you recorded months ago. Neither is wrong; it depends on how your table plays.

What are the four ways to transcribe a D&D session?

Broadly, there are four approaches. They trade off cost, effort, privacy, and how much the result understands that it is looking at a game.

1. A phone recorder and Whisper

Record the session on a phone or laptop, then run the file through Whisper, the open-source speech recognition model from OpenAI. You can run it on your own computer for free, or through one of the many apps and websites built on top of it.

Why you might choose it: it costs nothing but time, the audio never has to leave your computer if you run it locally, and the accuracy on clear English speech is excellent. Whisper also accepts an "initial prompt", a short piece of text that nudges it towards spellings you care about, which is a cheap way to help it with your campaign's names.

What to watch out for: plain Whisper does not know who is speaking. You get one long block of text with no names attached. Tools such as WhisperX add speaker diarization on top, but setting them up is a job for someone comfortable with Python and the command line. A four-hour session also takes a while to process on an ordinary laptop without a decent graphics card. And the result is only a transcript: summarising it, pulling out NPCs, or making it searchable is up to you.

Best for: technically inclined DMs who want full control, and groups where privacy matters more than convenience.

2. A general transcription service

Upload the recording to a general-purpose service built for meetings, interviews, and podcasts. Otter, Rev, and Descript are well-known examples, and there are many others.

Why you might choose it: it is easy. You upload a file and get back a clean transcript with speakers labelled, usually within minutes. Most offer a custom vocabulary list you can load with your campaign's names, and many produce an automatic summary too.

What to watch out for: these tools are built for business meetings, and it shows. Summaries tend to come out as "action items" and "key decisions," which is a strange way to describe a dragon fight. They have no idea that "Vex" and "Vex'ahlia" are the same character, or that the thing the party called "the creepy orb" in session two is the Orb of Kelvaran in session nine. Pricing is usually per minute or a monthly subscription sized for office use, and a weekly four-hour game uses up a lot of minutes. Read the privacy terms too: some services use uploaded audio to improve their models unless you opt out.

Best for: groups who want a readable transcript with minimum effort and do not need it tied into their campaign notes.

3. A Discord recording bot

If your group plays on Discord, a bot can join the voice channel and record for you. Craig is the long-standing favourite: it records each speaker to a separate audio track, which is exactly what a transcriber wants. Newer bots such as DiscMeet and Bardlog go further and produce a transcript and summary for you once the session ends, and Bardlog is designed specifically for D&D groups, with a campaign wiki built from each session.

Why you might choose it: nobody has to remember to start a recorder on their own machine, and per-speaker tracks mean speaker attribution is close to perfect. There is no hardware to buy.

What to watch out for: these bots are post-hoc. The transcript arrives after the session is over, so nobody can check what was said while you are still playing. Recording-only bots like Craig leave the transcription itself to you, so you are back to option 1 or 2 with better audio. Check how long each service keeps your recordings and who can download them; Craig, for example, deletes recordings after a set number of days, which is good for privacy but means you need to download them promptly.

Best for: online groups who want great audio with no effort, and are happy for the transcript to arrive after the game.

4. A tabletop-specific tool

The newest category is tools built specifically for tabletop RPGs. This is where Realms of Shod sits. Instead of treating the session as a meeting, they treat it as a chapter in an ongoing campaign: the transcript is split into scenes, each scene gets a short recap, and the NPCs, places, and items mentioned are linked into a realm compendium that grows every week.

Realms of Shod works two ways. Our Discord transcription bot joins your voice channel and transcribes live, so lines appear in the realm's Transcripts view as people speak. You can also upload a recording made any other way, including a phone on the table, and it goes through the same process afterwards.

Why you might choose it: the output is shaped for a D&D group rather than an office. You get scene recaps and a session summary, the transcript is searchable across the whole campaign, and a misheard name can be corrected once and fixed everywhere it appears.

What to watch out for: it is more than a transcriber, which means more to learn if all you wanted was a text file. AI features cost coins, so a long weekly campaign has an ongoing cost, although chat, dice, and the compendium are free. And like any tool that stores your game, it is only as private as its policies; ours are in the privacy policy, and recording is off until someone at the table turns it on.

Best for: groups running a long campaign who want notes they can actually use next week, not just a record of what was said.

Which one should you pick?

If you want it free and private and do not mind some setup, use Whisper. If you just want a transcript with minimal effort, a general service is fine. If you want perfect per-player audio from Discord, Craig is excellent at that one job. If you want the transcript to turn into campaign notes, recaps, and a searchable history, pick a tabletop-specific tool. Plenty of groups combine them: Craig for the raw audio archive, and something else for the notes.

Why does transcription get fantasy names wrong, and how do you fix it?

Every transcription tool will mangle your campaign's names at some point. Speech recognition models learn from huge amounts of ordinary speech, so when they hear an unfamiliar word they reach for the nearest familiar one. "Xanathar" becomes "Zanna Thar." Your player's carefully chosen elven name becomes a brand of yoghurt. The homebrew city of Vel'Sharra comes out three different ways in the same hour.

This matters more than it seems. A transcript you cannot search for a name is much less useful, because searching for names is the main thing people want to do with it. "When did we first meet Ilvara?" is unanswerable if she is spelt five ways.

There are three levels of fix, from least to most effort:

Tell the transcriber in advance. Whisper's initial prompt and the custom vocabulary lists in most commercial services both work this way. Before the session, give the tool a list of the names that are likely to come up. It helps a lot, but only for names you think of ahead of time, and every campaign introduces new ones mid-session.
Find and replace afterwards. Open the transcript in a text editor and fix the common mistakes. It works, and it is tedious, and you will miss the variants you did not think to search for.
Use aliases. Keep a list of the ways each name gets misheard or shortened, linked to the real one. In Realms of Shod, each Compendium entry keeps its own aliases: correct a misheard name in one line of the transcript and the tool offers to apply the same correction everywhere else it turned up. Nicknames work the same way, so "Tam," "Tamsin," and "that bard" all point to one NPC.

Whichever tool you use, the habit that pays off is the same. After a session, spend five minutes fixing the two or three names that came out worst. The fixes build up over the campaign, and the transcripts get better every week.

Do you need your group's permission to record?

Yes. Ask everyone at the table, every time someone new joins, and before you hit record.

Recording laws differ from place to place, and an online group can easily span several of them. We're not going to pretend to untangle that for you. You won't need to, though: if everyone has agreed, the question rarely comes up.

The bigger reason to ask is that recording changes how a table feels. Some players get self-conscious doing voices or playing a scene that matters to them when they know it's being kept. Some are fine with a transcript but not with audio sitting around. Some just want to know who can read it. None of that comes out unless you ask, and a player who feels recorded without a say is a player who quietly stops having fun.

A few habits that make the answer an easy yes:

Raise it at session zero, alongside lines and veils and the rest of the table's expectations. Say what gets recorded, where it goes, who can read it, and how long it's kept.
Make "no" a real option. If one player isn't comfortable, the answer for the whole table is no, or you find a version they're happy with, like a transcript with no audio kept.
Make pausing easy. Out-of-character chat about someone's week doesn't belong in the campaign record. Most tools can pause; on our Discord bot it's /pause and /resume.
Say it out loud when recording starts, especially when a guest joins. Many Discord bots announce themselves, but a spoken heads-up is still kinder.
Know how to delete it. Check that your tool lets you delete a session's audio and transcript, and do it if someone asks.
Think about who reads the recap. Sometimes a recap would spoil something for a character who wasn't there. Some tools, ours included, can keep the full recap to the players who attended and show everyone else a vaguer summary.

What does a realistic weekly transcription workflow look like?

The best workflow is the one your group will keep doing in week twenty, when the novelty has worn off. Here is a lightweight routine that works for most groups, whichever tool you use.

Before the session (two minutes):

Start the recording, or the bot, as soon as everyone is in the channel, and say so out loud.
If your tool supports it, add any new names you expect to come up tonight to its vocabulary or compendium.

During the session (no effort, ideally):

Pause for breaks and out-of-character tangents.
If your tool transcribes live, players can check what was said instead of asking the DM to repeat it.

After the session (five to ten minutes, the same night or the next day):

Stop the recording and make sure it saved. For recording-only bots, download the audio before it expires.
Skim the recap or summary. If you are transcribing yourself, this is when you run the file through your transcriber and write a short summary by hand.
Fix the two or three worst name mistakes.
Post the recap in your group chat. "Previously on..." is often the part of the whole process players enjoy most.

Before the next session (five minutes, for the DM):

Reread last session's recap, and search the transcript for any thread you want to pick up. This is where transcription earns its keep: prep based on what actually happened, not what you remember happening.

That is under twenty minutes a week in total, most of it optional. Compare that with the usual alternative, one player scribbling notes all session and missing half the game.

The short version

Good audio matters more than which tool you pick. Online, you mostly have it already; in person, move the recorder to the middle of the table and test it.
There are four realistic routes: Whisper for free and private, a general service for easy, a Discord recording bot for great audio, and a tabletop-specific tool for campaign notes.
Decide whether you want the transcript live, during the game, or afterwards. Most tools only do afterwards.
Every tool mishears fantasy names. Aliases, or a custom vocabulary, fix most of it.
Ask everyone before you record, make pausing easy, and know how to delete.

If you want to see the live approach in action, the Discord transcription bot page walks through setup in three steps, and the transcripts docs cover what you can do with a transcript once you have one.