Not a chatbot with a picture on it. A live webcam call: you talk, she hears you, answers in a voice you cloned, and her face moves as she speaks — thirty frames a second, uncensored, on hardware you control.
18+ only · Self-hosted models · Nothing leaves your infrastructure
Everything she needs, on one platform
Everything else on this page exists to serve this one moment: a person and their companion, live on webcam. This is a single turn of that call.
I missed you today. What are you doing right now?
Missing me was the best part of your day? Come here — tell me everything.
A depiction of one turn of the production pipeline — the same stages, in the same order, the live session runs.
Uncensored models on hardware you control. No provider policy deciding what your users are allowed to say or hear.
Face, voice and personality are bound to one character and held consistent across every session, image and frame.
Speech, reasoning, synthesis and image generation all run in containers you operate. No third-party AI provider ever receives your users’ audio, images or conversations.
A real-time, one-to-one webcam call with an uncensored companion. You speak into your mic; she listens, thinks, answers in her own voice, and her face moves on camera while she does it — every frame generated from the audio she is saying at that moment.
Your microphone is transcribed by a Whisper model running on your own hardware. The audio of an intimate conversation never reaches a third-party speech service, because there is no third party in the path.
Whisper
The transcript goes to a self-hosted language model with no content policy layered over it. She does not break character to decline, because there is nothing in the stack that can make her.
Self-hosted LLM
The reply is synthesised through the voice bound to that character — cloned from a single reference clip, or a preset. Word timings come back with the audio, so the mouth and the sound run off one clock.
GPT-SoVITS · Piper
The audio drives facial keypoints frame by frame and the result is pushed to the browser over WebRTC. Expression, blink and head pose are computed per frame — not a looping video, not a canned animation.
LivePortrait · 30 fps
The animation loop targets 30 fps for a live session. Every frame is generated from the audio she is speaking at that instant.
Per-frame keypointsA turn is a state machine, not a queue. Start talking over her and the current reply is cut off — the way a conversation actually works.
Interruptible turnsConversation memory persists per character. What she was told last week is in the prompt this week, so the relationship has continuity.
Persistent memoryAppearance is locked to the character record. She is the same person in session forty as in session one, and in every image you generate of her.
Appearance lockThe same pipeline takes a spoken sentence or a typed one. Text skips transcription and goes straight to the model.
Voice or text inputSessions publish into a WebRTC room. Create one server-side, hand the browser a short-lived token, and she appears on your site instead of ours.
LiveKit · embeddableEvery major hosted AI provider forbids adult content in its own terms of service. That is the moat — and it is the reason this runs on models nobody can revoke.
| Building on a hosted provider | Building here | |
|---|---|---|
| Adult content | Forbidden by the terms of every major hosted AI provider | The category the platform is built for |
| Refusals mid-conversation | A filtered model breaks character to decline | Uncensored weights — there is no refusal layer |
| Your users’ audio and images | Uploaded to a vendor who may retain or review it | Never leaves the containers you operate |
| Policy risk | One terms update can end your product | No provider policy sits above yours |
| A face that responds live | Text, or a pre-rendered clip | Keypoints generated per frame at 30 fps |
| Cost as you grow | Metered per token, scaling with engagement | A hardware question, not a per-call bill |
This is the same synthesis engine a session uses. Type anything; the waveform is decoded from the file you are about to hear.
Zero-shot cloning builds a persistent voice from a short sample. Preset neural voices cover everything else, streaming with word-level timestamps so lips and subtitles run off the same clock.
Generate a character from a prompt or a reference, then hold that identity across every later image and frame. Appearance, personality and voice travel together as one character.
Self-hosted language models with no content filter and no provider terms to comply with. Memory persists between sessions, so she remembers what she was told last time.
Audio-driven lip-sync video for longer scenes, pose-driven full-body animation, and text or image to video for anything pre-rendered.
Each runs on a model we host. Where something is not finished yet, it says so.
Embed a live session with a few lines. REST and WebRTC, server-side keys with scopes and quotas, and short-lived session tokens for the browser so a key never reaches a client.
Read the API docsconst session = await arapulse.sessions.create({
characterId: "rissa",
voice: "cloned:rissa-v2",
realtime: true // WebRTC video + audio
});
// Embedded sessions default to 12 fps; pass target_fps to raise it.
// The portal's own live session runs at 30.
session.on("frame", render);
session.say("Hey, you're back.");