Kitt
KITT is an AI tool that allows users to have live conversations with ChatGPT using WebRTC. This tool supports multiple use cases, such as v...
Last verified:
What is Kitt?
KITT is an AI voice assistant that joins WebRTC conference calls, powered by ChatGPT, Google Cloud Speech-to-Text, and Text-to-Speech. Users can have live, real-time conversations with it in video meetings, where it answers questions, takes notes, summarizes discussions, and translates languages. Built using LiveKit's real-time audio/video framework, KITT automatically joins sessions via webhooks and subscribes to participants' audio tracks for seamless interaction.
Key features include low-latency streaming STT with 20ms frames for quick transcription, streaming GPT-3.5 Turbo responses delimited by sentences for natural speech flow, and real-time TTS to broadcast responses to all users. Activation occurs by saying 'KITT' in group calls, with live transcription display and animated visual states. The thin client uses LiveKit Meet UI, keeping the heavy processing server-side.
It is designed for developers building AI companions in video apps, teams wanting smart meeting assistants, and anyone experimenting with multimodal AI in WebRTC environments. KITT demonstrates production-grade latency optimization and modular architecture for easy extension.
Kitt pricing
Pricing model: Paid
KITT demo at https://livekit.io/kitt is free to use. Self-hosting requires LiveKit Cloud account (free tier with limits, paid for production scale), OpenAI API usage (pay-per-token for GPT), and Google Cloud billing for STT/TTS (no free tier for streaming, charged per minute/character). No dedicated pricing plans; costs depend on usage.
Kitt pros
- Real-time voice conversations with ChatGPT
- Automatic session joining via webhooks
- Low-latency 20ms STT frames
- Streaming GPT responses by sentence
- Multi-language support and translation
- Meeting note-taking and summarization
- Live transcription display
- Animated visual states for engagement
- Group call activation with 'KITT' keyword
- Server-side processing for thin client
- Subscribes to all participant audio
- Uses reliable Google STT and TTS
- Open-source code on GitHub
- Optimized for human-like latency
- Contextual conversation history
- Works in LiveKit Meet UI
- Supports 1:1 and multi-user calls
Kitt cons
- Requires Google Cloud credentials
- Needs OpenAI API key
- Depends on LiveKit Cloud account
- Original repo archived since 2024
- Uses paid GPT-3.5 Turbo model
- Google STT incurs costs
- Limited to microphone audio input
- No video processing yet
- Robotic-sounding TTS voice
- English-centric prompt tuning
- Local setup needs manual API call
- No on-device STT fallback
Frequently asked questions about Kitt
How does KITT join a meeting?
When a new session starts and the first user joins, a LiveKit webhook triggers the lkgpt-service to connect KITT as a participant using the Go SDK, subscribing to all microphone tracks.
What STT service powers KITT?
Google Cloud Speech-to-Text with streaming recognition at 20ms frame size for low latency, chosen over Deepgram and Whisper for speed and accuracy balance.
How is GPT prompted for concise responses?
System prompt instructs 'You are a voice assistant in a meeting named KITT, make concise/short answers,' with full conversation history and language codes prepended.
What activates KITT in group calls?
Saying 'KITT' or 'Hey KITT' signals the prompt is for it, playing a listening sound; in 1:1 calls, all speech is directed to KITT.
How are GPT responses turned into speech?
GPT-3.5 Turbo streams tokens, delimited into sentences; each sentence is synthesized via Google TTS and queued to audio track for real-time broadcast.
What client UI does KITT use?
LiveKit Meet, a Zoom-inspired NextJS video call app, showing live transcription, KITT's animated states via React, and data messages for sync.
Is the KITT repo still maintained?
Original repo archived Jan 2024; KITT V2 uses new Agents Framework, accessible via LiveKit Playground at kitt.livekit.io.
What are local setup requirements?
Run meet (Node.js) and lkgpt-service (Go), configure env with LiveKit/OpenAI/Google creds, manually POST to /join/room_name for local webhook simulation.
Can KITT handle multiple languages?
Users specify language; prompts include codes like en-US, enabling responses in native tongues and live translation between speakers.
What optimizations reduce latency?
Streaming everything: incremental STT, sentence-delimited GPT, real-time TTS; GPT primed for brevity; no full-response blocking.