Projects
Everything is on GitHub and Hugging Face.
- Benchmark
OpenVideoFDB
An open, runnable implementation of VideoFDB, a benchmark for full-duplex audio-visual conversational agents. A real-time harness that streams a clip into a live agent and records what it says back, the paper’s timing metrics and LLM-judge rubrics, adapters for Gemini Live, OpenAI Realtime and three avatar vendors, and an open clip set mined from Meta’s Seamless Interaction corpus for anyone without access to the gated NVIDIA data.
- Model
smart-turn-streaming
Silence-aware streaming end-of-turn detection for voice agents, built on Pipecat’s Smart Turn v3.2 encoder. Scores the conversation every 100 ms instead of once, decides at 0.2, 0.5 and 1.2 seconds of silence with thresholds fitted to a false-alarm budget, and drives Pipecat’s speculation path so the language model starts answering before the turn is committed. CPU only, 32 MB of ONNX model files.