VocRT: Real-Time Conversational AI built entirely with local processing (Whisper STT, Kokoro TTS, Qdrant)

Anurag Singh — Mon, 30 Jun 2025 21:38:13 GMT

I've recently built and released VocRT, a fully open-source, privacy-first voice-to-voice AI platform focused on real-time conversational interactions. The project emphasizes entirely local processing with zero external API dependencies, aiming to deliver natural, human-like dialogues.

Technical Highlights:

Real-Time Voice Processing: Built with a highly efficient non-blocking pipeline for ultra-low latency voice interactions.
Local Speech-to-Text (STT): Utilizes the open-source Whisper model locally, removing reliance on third-party APIs.
Speech Synthesis (TTS): Integrated Kokoro TTS for natural, human-like speech generation directly on-device.
Voice Activity Detection (VAD): Leveraged Silero VAD for accurate real-time voice detection and smoother conversational flow.
Advanced Retrieval-Augmented Generation (RAG): Integrated Qdrant Vector DB for seamless context-aware conversations, capable of managing millions of embeddings.

Stack:

Python (backend, ML integrations)
ReactJS (frontend interface)
Whisper (STT), Kokoro (TTS), Silero (VAD)
Qdrant Vector Database

Real-world Applications:

Accessible voice interfaces
Context-aware chatbots and virtual agents
Interactive voice-driven educational tools
Secure voice-based healthcare applications

GitHub and Documentation:

Code & Model Details: VocRT on Hugging Face

I’m actively looking for feedback, suggestions, or potential collaborations from the developer community. Contributions and ideas on further optimizing and expanding the project's capabilities are highly welcome.

Thanks, and looking forward to your thoughts and questions!