GABRIEL

ready

Live transcription

press REC and speak…

idle

Self-talk reasoning

step-by-step · same sequence

idle

LIVE process flow — every step lights while it runs on ANY tab or device (server-driven), visualizer trace-mode style.

🎤 MIC

device microphone · mob/pc

VAD

voice gate · chunker

STT

—

SPEAKER ID

voice match · runs with STT

THINK

—

TTS

—

🔊 OUT

audio playback

HEARD (STT)

—

REPLY (THINKING)

—

VOICE (TTS)

—

TRACE — server pipeline timeline (Δms between stages)

—

total latency—

Activity feed

stage transitions appear here live…

Active pipeline (tap a slot → highlight → pick ANY model to substitute)

TTS ▲—

THINKING—

STT ▼—

Open chat context usage —

Context cap: 4096 tokens (model max: —)

Models (type = instant local filter · GO = search Ollama + HuggingFace)

Thinking models (LLM)

STT models (speech → text)

TTS models (text → speech)

Generated voice storage

WHO IS TALKING

Each profile = a person. Enroll 3+ voice samples per person. Gabriel recognizes them automatically and adapts its response voice.

GPU

loading…

CPU / RAM / DISK

loading…

Server log

System prompt

Voice

Speaking language Accent / dialect Voice type Tone Auto-interject in group chats — Gabriel jumps into member-to-member talk (Sarcastic = dry digs, Curious = connects the dots), up to 3 quick messages Voice speed: 1.00x

Speaker recognition

Lower = merges more aggressively (one person stays one profile). Higher = stricter (may split a person into many). Defaults: match 55, cluster 50.

Match threshold (enrolled profiles): 55

Cluster threshold (auto-discovery): 50

Server

Version

App —

Server —

Password

Current password New password (min 4)

Notifications

—

Server

Permanent memory accumulated from chats & live (RAG). Editable. Facts inject into future conversations.

load to view…