Give your NPCs an AI voice in Foundry VTT.

What it does

Voice turns the AI's words into speech. When a character talks or you ask for narration, Familiar sends the line to the text-to-speech provider you picked and plays the audio back in the chat, right there in Foundry. The table hears the scene instead of reading it off the screen.

You set two voices. A DM or narrator voice carries description and the beats between people; a character voice carries the people themselves. The AI chooses between them line by line, so a stretch of narration and the innkeeper answering it come out in the right voice without you switching anything.

• Two voices: a DM or narrator voice and a separate character voice, picked per line by the AI.
• Several characters in one exchange play back in order, gapless, as a single take.
• Speaks the narration and the action too, so the parts between the quoted lines are heard, not skipped.

Your key stays in your browser

Voice is bring-your-own-key. You paste your own provider key into Familiar's voice settings, and the audio is made browser-direct: your Foundry tab calls the provider, decodes the reply, and plays it. The key lives in your browser. It never reaches a Familiar server, because there is no Familiar server in the audio path.

So there is no extra subscription to us and no second bill from us. You pay your provider for what you generate, the same way the rest of Familiar handles keys, and nothing routes through anyone else on the way.

Nothing sits between you and the provider. The key you paste stays on your machine, and the audio is generated and played in your own browser.

Set it up

Four steps from nothing to a talking NPC. You do this once.

1. 1

   Open voice settings

   From Familiar's chat window in Foundry, open the voice settings.

2. 2

   Paste a provider key

   Pick a provider and paste your API key. The key is per-user and stays in your browser.

3. 3

   Choose a voice and test

   Choose a voice, optionally pick a model, and hit Test to hear it. Tune it until the voice sounds right.

4. 4

   Add a character voice (optional)

   Set a second voice in the Character slot if you want narration and the cast to sound different. Leave it blank and every line uses your first voice.

Pick a provider

Every provider here is bring-your-own-key and browser-direct. They trade cost against richness, so pick by how your night runs and what you want to spend. The figures below assume a typical session, around 15,000 characters of speech, and you pay the provider directly.

• ElevenLabs has the largest voice library and a fast multilingual default. Pick the eleven_v3 model on a voice for audio tags and singing.
• Cartesia is sub-100ms and covers 42 languages, on the sonic-3.5 model. Its voices are UUID ids, so the in-app dropdown loads them after you paste your key.
• OpenAI reuses the OpenAI key you may already have. Models run from gpt-4o-mini-tts, which you can steer, through tts-1 to tts-1-hd for the highest quality, with named voices like alloy and nova.
• Other routes sit in the same settings. An OpenRouter key reaches TTS models from several labs on one bill, a NanoGPT balance covers speech alongside its chat and image models, and a custom entry points at any OpenAI-compatible voice server you run. None of the three ships a voice list, so you type the voice name yourself. OpenRouter and NanoGPT do load their model lists for you; on a custom server you name the model too.

A simple split: ElevenLabs for a narration-heavy night, Cartesia when cost matters, OpenAI as the cheap default if you already hold an OpenAI key.

Give each NPC its own voice

Past the two slots, you can pin a voice to one character. Assign a voice to an NPC once and the AI uses it automatically every time that NPC speaks, so the harbourmaster sounds like the harbourmaster from one session to the next.

This is the audible half of writing a character down. A written anchor keeps an NPC consistent on the page; a pinned voice keeps them consistent in the ear. The companion guide on writing characters covers the anchor side.

• Voice & Image Generation

Emotion and singing

An optional emotion hint colours how a line is delivered. Tag it and the voice leans into the cue, from a whisper to a shout; leave it off and the line is read plainly.

Provider support differs. Cartesia and OpenAI take an emotion hint on any voice. ElevenLabs renders it only on the eleven_v3 model, the same model that unlocks singing, and skips it on the faster default.

Singing is the special case. Set the emotion to singing and the line is sung rather than spoken. It works on an ElevenLabs eleven_v3 voice only, so put that model on the voice you want to carry a song.

• Emotion hints: angry, sad, happy, excited, calm, scared, whispering, shouting, nervous, tired, and singing.

On ElevenLabs, switch the voice to the eleven_v3 model for emotion or singing; the faster default skips both. Cartesia and OpenAI take an emotion hint as it is.

If a voice will not play

One thing trips people up after switching providers: a voice has to belong to the provider it is set on. An OpenAI voice name like alloy will not play on Cartesia, which identifies its voices by UUID rather than by name.

The dropdown is the fix. After you paste a provider key, choose a voice from the in-app list, which loads that provider's full library. Picking from the dropdown rather than carrying a name over from another provider keeps the voice and the provider matched.

A saved voice the provider no longer has stays in the list, marked unavailable, instead of quietly playing something else. That label is the fastest way to spot an id carried over from an older setup.

Voice one NPC and play

Pick a character who matters to your next session, set a provider and a voice, and let the AI speak the next time the party talks to them. You stay in the scene while it carries the lines. Questions about voices, providers, or your first session are welcome in the Discord.

Looking for session transcription instead, speaking your table aloud and saving a searchable record? That is a separate feature, with its own guide on the way.