When the same innkeeper speaks at the beginning of session two and session twelve, the table should recognize them before the name appears on a map. That does not require a perfect character performance. It requires a few choices that stay put.
AI voice for tabletop RPG campaigns is most useful when it creates small, reusable artifacts around play: the cold-open recap, the warning sealed in a letter, the recorded testimony, or the returning NPC who has one familiar cadence. In Vois, keep those scripts, speaker assignments, pronunciation notes, and generated takes inside the campaign project so the GM can revise an asset without recreating the cast. The GM still reads the room, improvises, and decides when silence is better.
Build a recurring NPC voice bible first
A voice bible is a compact record of how a character should sound and when that audio should be used. It is not a casting sheet. Keep it short enough to update during campaign prep.
Start with the NPCs who return. A guard who speaks once can stay as text at the table. The smuggler who offers information every few sessions deserves a stable entry.
| Field | What to record | Example |
|---|---|---|
| Character ID | A stable short label | NPC-MARIN |
| Role at the table | Why the character returns | Harbor contact |
| Voice cues | A few plain-language choices | Low energy, measured pace, dry humor |
| Guardrails | What not to do | No imitation of a real person; no exaggerated accent |
| Name notes | Approved spoken form | Ma-rin, stress on first syllable |
| Test line | One line you listen to before a batch | "You arrived late, which means the tide is on our side." |
Use a separate entry for a narrator or a recurring document reader. If every role gets the same serious, theatrical delivery, the campaign starts to sound like a trailer. Contrast comes from pacing, word choice, and role, not from pushing voices into caricature.
When you are ready to assign the cast, test the approved line in Vois's voice library. Save the chosen voice in the bible and use it for every return appearance. A consistent cast usually starts with a small set of recurring roles.
Make a few campaign assets do more work
Tabletop audio earns its place when it gives players something they can hear again or react to together. Build a short library rather than a continuous soundtrack.
Good first assets include:
- a one-minute recap voiced by the campaign narrator;
- an NPC's letter or journal excerpt, recorded as a prop the group can replay;
- a dispatch, proclamation, or witness statement that contains a clue;
- a scene opener that states the place, time, and immediate tension; and
- a compact farewell or warning for a character who may return later.
Write recaps as memory prompts, not a transcript of the previous session. Name the unresolved decision, the person involved, and the physical situation the group returns to. Leave out tactical detail that the players need to discover again.
For prop audio, give the speaker a reason to be recorded. A captain dictated a message because the ship was leaving. A librarian made a private note because the archive could not be removed. This small bit of fiction tells players why they are listening and makes the audio feel like an object in the world.
Pair multilingual names with pronunciation notes
Fantasy worlds often mix invented names with real languages or language-inspired terms. Treat those terms carefully. A phonetic spelling in the voice bible is more useful than a note that says "say this elegantly."
The pronunciation dictionary lets you store a written form, a phonetic form, and a context note. Keep campaign-specific names in the project scope so they do not spill into another story. Test each recurring name in a full line, because a name that works by itself can sound different between two other words.
If a letter or scene includes dialogue in another language, get the text reviewed by someone fluent before generating it. Do not invent a translation from a sound you like. W3C's language-tag guidance and the linked IANA registry are useful references when you need to label language-specific source material in your campaign files. Vois Pro includes support marketed as 600+ languages, but a long language list does not replace a human check of a name, phrase, or cultural reference.
Get clear consent and skip celebrity impersonation
A cloned voice should begin with a clear conversation, not a surprise at the table. Ask the voice owner whether the clone may be used for the current campaign, session recaps, recorded props, or a public actual-play recording. Put the answer and the expected retention period in the voice bible. If they later change their mind, stop using the clone and remove it from upcoming material.
Vois voice cloning needs a clear 10-15 second audio sample and the voice owner's permission. Record a fresh sample when possible. Do not pull an old group recording apart to create a clone without asking everyone involved.
And do not ask for a celebrity or public figure imitation. "Sound exactly like [person]" is not a character direction. Describe an original voice instead: careful diction, a tired tempo, a formal greeting, clipped answers, a soft laugh. This produces a character your table can own without borrowing someone else's identity.
The U.S. Copyright Office's report on digital replicas is useful background for anyone thinking seriously about recorded voices. It is not a replacement for direct permission from the person in your group.
Generate a session packet in batches
Do the writing and casting before the generation pass. Create one Vois project for the campaign, then make a script for each asset instead of collecting loose clips on the desktop.
A simple naming pattern is enough:
ARC-01-S01-recap
ARC-01-S01-letter-from-marin
ARC-01-S01-dockside-opening
ARC-01-S01-marin-entrance
For a dialogue-heavy opener, make each role a distinct speaker in the project. The labels below show the finished scene, not a command to run:
Narrator: Rain moves across the harbor before anyone speaks.
Marin: You arrived late, which means the tide is on our side.
Prompt your agent: "Create a session-one audio plan from this NPC voice bible and these approved scene notes. Separate recaps, letters, scene openers, and optional NPC entrances. Preserve the supplied speaker labels and pronunciations, do not imitate real people, and return the proposed asset list and cast for GM approval before generation."
Review the plan in this order:
- Confirm that every line advances a scene or gives players a replayable clue.
- Check character names and voice directions against the voice bible.
- Approve the exact cast and script in Vois, then generate a single representative asset.
- Listen at the table setup, adjust only approved pacing or pronunciation changes, and generate the remaining approved assets.
Use Multi-Speaker when a reviewed scene needs its cast rendered together. Keep player prompts in your GM notes unless you deliberately want a narrator to read them aloud, so the clip stops before the table needs to act.
Organize playback around the session, not the character list
A folder with one hundred NPC names is not helpful when the group has just opened a suspicious letter. Organize playback around the point of use.
Campaign-Aster/
S01-The-Quiet-Harbor/
01-recap/
02-letters-and-handouts/
03-scene-openers/
04-optional-npc-entrances/
cue-list.txt
Give every clip a spoken name in its filename, followed by a brief cue. 02-letter-marin-play-after-map.mp3 is easier to use than final-v4.mp3. In cue-list.txt, write the trigger in one sentence: "Play after the group accepts the chart" or "Skip if they leave the harbor."
Keep optional clips separate from the must-play folder. An RPG session can take an unexpected turn in two minutes. You want the table to feel supported by the audio, not held hostage by your prep.
Let audio make room for play
Start with one recap, one recurring NPC, and one letter. If the group leans in, add another artifact next session. If they are already talking over the clip, make the next one shorter or turn it into a handout.
Your campaign does not need a full cast to feel alive. It needs a few voices that arrive at the right moment. Get started with Multi-Speaker, then download Vois and give the players one thing they will remember hearing.
Sources
- U.S. Copyright Office, Copyright and Artificial Intelligence, Part 1: Digital Replicas
- W3C, "Language tags in HTML and XML"
Leave space for the players' choices. The most memorable line is often the one no one prepared.
The Vois Team