Tutur is a privacy-first, zero-latency AI speech recasting web application designed for non-native English learners. Built to help users practice spoken English without fear or judgment, Tutur listens to incoming real-time audio streams, transcribes spoken phrases, and instantly reformulates them into natural, fluent native expressions while preserving the user's original intent and tone.
Powered by a FastAPI backend integrated with FastRTC for WebRTC audio handling, the system utilizes Silero VAD for intelligent pause detection. Spoken audio is processed on-the-fly via Groq's high-speed API—leveraging Whisper Large v3 Turbo for Speech-to-Text and GPT OSS 20B for real-time sentence recasting. The frontend is built as a reactive Single Page Application (SPA) using Alpine.js and Tailwind CSS, communicating over WebRTC data channels for instant live feed updates.
| Screen | Description | Preview |
|---|---|---|
| Idle Companion Workspace | The minimalist initial state displaying session status indicators, microphone trigger, and usage guidance. |  |
| Live Recast Session | Real-time interactive session feed showing speech-to-text transcriptions ("You") alongside native reformulations ("Versi natural") as you speak. |  |
- Real-Time WebRTC Audio Streaming: High-throughput, send-only audio streaming over WebRTC paired with server-side Silero Voice Activity Detection (VAD) that triggers recasting upon natural pauses.
- Dual-Stage Groq AI Pipeline: Ultra-fast Speech-to-Text using
whisper-large-v3-turbocombined withopenai/gpt-oss-20bfor instant grammar, tense, and collocation reformulations. - Dynamic ICE/TURN Credential Management: Automated background renewal and serving of short-lived TURN/STUN credentials to guarantee seamless WebRTC NAT/firewall traversal.
- Zero-Storage & Privacy-First: Audio buffer arrays are converted to PCM in-memory on demand; no raw audio files, transcriptions, or user sessions are ever stored on disk.
- Reactive Stream UI: Lightweight Alpine.js SPA with auto-scrolling conversation feed, connection state management, and real-time visual distinction for already-natural phrasing.
- Backend Framework: FastAPI (Python)
- Real-Time Streaming Engine: FastRTC (
fastrtc) over WebRTC - Voice Activity Detection: Silero VAD (
SileroVadOptions&ReplyOnPause) - Speech-to-Text (STT): Whisper Large v3 Turbo (
whisper-large-v3-turbovia Groq Async API) - Language Model (LLM): GPT OSS 20B (
openai/gpt-oss-20bvia Groq) - Network Traversal: Dynamic STUN/TURN Relay Server Management
- Frontend Architecture: Single Page Application (Alpine.js / Tailwind CSS / FontAwesome)
- Audio Processing Engine: NumPy, Wave, In-Memory BytesIO
- License: MIT License
This project is licensed under the MIT License - see the LICENSE file for details.