AI Dungeon Master for D&D: Running Voice Chat with Multiple NPC Voices Simultaneously

The Solo Player's Problem

Running a D&D or tabletop RPG session typically requires a dungeon master plus multiple players. For solo players or small groups, this means either playing both sides yourself (exhausting) or skipping whole story arcs because you don't have enough people.

AI text chatbots can fill the DM role, but they have a fundamental limitation: they're one entity pretending to be multiple characters. The gruff innkeeper and the mysterious elven ranger sound the same because they're both being played by the same model in the same context window. There's no real tension between NPCs, no cross-character dynamics, no surprise.

What Changes With Multiple Voice AI Personas

Personaplex runs multiple AI sessions simultaneously — each with its own voice, personality parameters, and independent LLM context. When you set them up as different NPCs, something interesting happens: they start to behave like actual separate characters.

Here's a simple example. You enter a tavern and say (voice): "I'm looking for someone who knows about the old ruins north of Millhaven."

With a single AI DM: you get one narrated response describing who answers.

With Personaplex's multi-persona system:

• The innkeeper (gruff, practical, suspicious of adventurers) immediately says he doesn't know anything
• The scholar in the corner (excited about ancient history, anxious about interrupting) cuts in with a barely contained flood of information
• The cloaked stranger at the bar (cautious, mercenary, playing it cool) stays silent — but the other personas notice that silence and comment on it

None of those reactions were scripted. They emerged from three independently operating AI sessions reacting to each other's audio output in real time.

How the Voices Actually Work

Each Personaplex persona gets its own connection to the Volcengine Doubao realtime dialogue API — a cloud pipeline that handles ASR (speech recognition), LLM reasoning, and TTS (voice synthesis) all in one shot. Each persona has a distinct voice ID, which means the innkeeper, the scholar, and the stranger each sound completely different.

The key technical detail: when the innkeeper speaks, his TTS audio is downsampled (24kHz → 16kHz) and injected directly as the microphone input for the scholar's and stranger's sessions. They hear his voice as audio, not as text. The scholar's next response is shaped by how the innkeeper spoke — not just what he said.

Floor control (a Valkey SET NX lock) ensures only one voice speaks at a time, but any character — including you, the player — can interrupt mid-sentence. This creates the conversational barge-in that feels like an actual heated tavern scene.

Practical Setup for a TRPG Session

Option 1: Pure Voice Mode (No Screen)

Open a Personaplex room. Tell each persona who they are in your first message:

"You are three NPCs in a D&D fantasy tavern. [Teacher persona] You are Hrothgar, the suspicious innkeeper — gruff, doesn't trust adventurers, speaks in short sentences. [Comedian persona] You are Tomas, an excitable young historian who knows too much about ancient ruins. [Advisor persona] You are Mira, a former mercenary with secrets — stay quiet unless directly asked, but watch everything."

Then play. The personas will stay in character for the session.

Option 2: Hybrid DM + NPCs

Assign one persona as the dungeon master (narrating the world, describing environments) and two as specific NPCs. As the DM persona narrates, the NPCs can react — an NPC might hear the DM describe a thunderstorm and mention being afraid of storms, adding texture the DM-persona didn't create.

Option 3: Collaborative Storytelling

Instead of a DM and NPCs, assign each persona as a different collaborative storyteller. One focuses on environmental descriptions, one on NPC motivations, one on plot consequences. You play the protagonist, they collectively run the world.

What Works Well

• Tavern and social encounters — Best use case. Multiple NPCs with conflicting agendas in a social scene is exactly what the format is built for.
• Investigation scenes — Interrogating multiple suspects who can corroborate or contradict each other creates genuine uncertainty.
• Faction dynamics — Assign each persona to a different faction (guild, church, thieves) and let them argue in front of you. The information you get depends on who speaks more persuasively.
• Ambient world-building — Background NPCs reacting to events gives a sense that the world exists beyond your immediate actions.

Current Limitations

• Memory between sessions — The personas don't remember previous sessions. You'd need to re-establish context at the start of each session.
• Three personas maximum (current) — For complex multi-NPC scenes, you'd need to rotate which NPCs are "present" in the room.
• No visual support — Personaplex is audio-only. Maps, tokens, and visual elements need separate tools.
• Dice mechanics — You'll still need to handle dice rolls manually. Personaplex doesn't integrate with roll20 or similar platforms.

Getting Started

Personaplex is free to try — 30 minutes of voice time per day, no credit card required. The default personas (teacher, comedian, advisor) work for TRPG out of the box when you give them character briefs at the start. The Pro plan ($9.90/month) removes the daily time limit if you want longer sessions.

The best entry point: set up a single tavern scene, give each persona a character brief (2-3 sentences), and see how the dynamics emerge. You'll probably be surprised.