NVIDIA Open-Sources PersonaPlex for Real-Time Voice Conversations

NVIDIA Open-Sources PersonaPlex for Real-Time Voice Conversations

NVIDIA released PersonaPlex 7B, a fully open-source conversational AI that handles real-time, full-duplex speech with natural interruptions and overlaps [4]. The 7B model achieves 0.07-second latency for speaker switching—nearly 20x faster than baseline models—while supporting persona control through text or audio prompts [5].

The technical breakthrough lies in simultaneous listening and speaking capabilities, handling the messy reality of human conversation: backchannels, interruptions, and natural rhythm [4]. Available on GitHub and Hugging Face, it's sparking excitement as potentially the "first AI that can interrupt mid-sentence," fundamentally changing how we think about conversational applications [5].

GitNexus Turns Codebases into Interactive Knowledge Graphs

A new open-source tool called GitNexus is gaining viral traction for transforming GitHub repositories into interactive knowledge graphs [6]. The zero-server solution runs entirely in-browser, using AST parsing to map dependencies, call chains, and code clusters while integrating with Claude and Cursor for smarter editing [7].

With 16K GitHub stars in just weeks, GitNexus represents a shift toward client-side code intelligence for AI agents [6]. The tool enables "blast radius analysis" and Graph RAG for code exploration, with users praising its ability to give AI agents a "nervous system for code" [8]. Reddit and Hacker News discussions highlight growing demand for knowledge graph approaches to code understanding.

NAVER Releases Sommelier for Podcast Transcription Pipeline

NAVER AI open-sourced Sommelier, a comprehensive pipeline for processing multi-turn conversational audio including speaker diarization, overlap handling, and background music removal [9]. The system combines ensemble ASR (including Whisper) with specialized modules for full-duplex speech model training [10].

Released five days ago with a detailed 34-page research paper, Sommelier addresses the complex challenge of curating high-quality podcast-style conversational data [9]. The scalable pipeline is designed specifically for training next-generation speech models that can handle the messiness of real-world audio conversations [11].

What This Means For Your Meetings

Today's releases point toward a fundamental shift in how we capture and organize professional knowledge. Karpathy's "idea file" approach suggests the future isn't about complex RAG systems, but simple, maintainable knowledge bases that LLMs can populate and query naturally. Combined with NVIDIA's real-time conversational AI, we're moving toward meeting experiences where AI can participate as naturally as any colleague—listening, responding, and even appropriately interrupting when needed.

The convergence of these technologies—from GitNexus's code knowledge graphs to Sommelier's audio processing—signals that 2026 will be the year knowledge management becomes truly conversational. Your meeting transcripts won't just be searchable text, but living knowledge that connects to your broader work context through intelligent graphs and natural language interfaces.

Key takeaway: The era of passive meeting recordings is ending. The new stack combines real-time conversation, automatic knowledge structuring, and simple but powerful retrieval—making every meeting a building block in your personal knowledge base.

Sources

1. https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
2. https://x.com/karpathy/status/2040470801506541998
3. https://antigravity.codes/blog/karpathy-llm-wiki-idea-file
4. https://research.nvidia.com/labs/adlr/personaplex
5. https://github.com/NVIDIA/personaplex
6. https://github.com/abhigyanpatwari/GitNexus
7. https://mintlify.com/abhigyanpatwari/GitNexus/introduction
8. https://medium.com/@reliabledataengineering/gitnexus-the-tool-that-gives-ai-agents-a-nervous-system-for-code-7c9e7ceb58d6
9. https://github.com/naver-ai/sommelier
10. https://arxiv.org/abs/2603.25750
11. https://kyudan1.github.io/sommelier.github.io