Morphius — Reactive AI Face & Voice Assistant

SYSTEM METRICS

AMPLITUDE --

PEAK FREQ --

ENERGY --

STABILITY --

STATE IDLE

CONSOLE

Voice ON

LLM Bridge // Settings

Provider

Model Name Example: gemini-2.5-flash or llama3.2

WebLLM Model Runs fully in-browser. The model downloads once, then loads from cache. Bigger models need more RAM/VRAM.

API Key Required for cloud chat providers (Gemini, OpenAI, Claude, ...). Leave empty for Local Agent / WebLLM / Ollama.

Base URL / Endpoint Auto-configured per provider

Temperature

Max Tokens

Voice Language (TTS + STT) Used for both speech output and microphone recognition

Auto-send after voice input

Master Volume 100%

Output Device Playback routing where the browser allows it. Browser voices always use the system default device.

Microphone Engine On-device = in-browser Whisper. First use downloads a small model, then it works offline. Audio never leaves your machine.

Whisper Model Size Bigger models transcribe more accurately but download and run slower on-device.

Speech Output Engine On-device TTS: Kokoro is excellent for English; Arabic uses Meta MMS (more robotic than browser Arabic voices).

Use Gemini native voice (Text-to-Speech) Handles audio I/O and speech generation natively -- more reliable across languages.

Live voice call (Gemini Live API -- WebSocket) Real-time two-way voice chat: mic stays open and streams audio directly via WebSocket.

Live face mirror (on-device) MediaPipe FaceLandmarker reads your face from the camera. Frames are analyzed on your GPU locally -- nothing leaves this device.

MIRROR: OFF

Sentiment source Detects the feeling of the words being typed or spoken (English + Arabic) and colors the face accordingly. Louder speech adds a warm glow on top.

Real-time object detection (AI Vision) Detects objects from your camera on-device (YOLO26n via WebGPU, or COCO-SSD fallback). Frames stay on your device.

Vision Backend Auto picks WebGPU YOLO when available, otherwise falls back to COCO-SSD on WebGL.

Max detection rate (FPS)

React when a person is detected Nudges the avatar into a subtle engaged look while a person is in view.

Speak about what I see When idle, the avatar may comment out loud on what it detects — people, animals, new objects.

VISION: OFF

Behavioral intelligence (on-device) Classifies your interaction (help, confusion, positive, negative, …) and adapts the avatar's reaction and response-cadence hints locally. All data stays on this device. Turn off to restore the original behavior exactly.

System Prompt

3D Model Switch to Custom to load your own GLB, GLTF or VRM model.

Load Model

No custom model loaded

Show 3D Face Show/hide the 3D face model. Uncheck to use only visual effects.

Visual Style Audio-reactive visual effect behind the face.

Viseme Playground

Press the letter, then hold Preview to make the face mouth that shape: M, O, E, A, S, T, L, K, C, F, U.

Auto-restore last chat On load, resume your most recent conversation.

Theme

Lighting Preset Global ambient light tint and brightness.

Interface Language

AI Face v7.3.0 — Offline AI Face & Voice Assistant

Loading model 0%

Initializing engine...

SYS.NODE // 0x0000

00:00:00

X:000 Y:000

0.00 Hz

STATE // IDLE

API: window.AIFace -- see console

Listening...

Sessions

Conversations are saved automatically. Pick one to resume.

Map Morph Targets

Match your model's morph names to the avatar's canonical slots so expressions, lip-sync and the mirror work.

Sessions

Map Morph Targets

Welcome to AI Face