SoulX-Duplug: Plug-and-Play Streaming State Prediction Module for Realtime Full-Duplex Speech Conversation

A large research team from academia has introduced SoulX-Duplug, a novel open-source module designed to solve the persistent challenges in creating human-like, real-time voice AI. The system tackles issues like catastrophic forgetting (where AI loses old skills when learning new ones), scarce training data, and poor scalability by offering a 'plug-and-play' component that can be integrated into existing spoken dialogue systems. Its core innovation is performing streaming Automatic Speech Recognition (ASR) and then using the resulting text in real-time to predict the user's intent and the overall state of the conversation. This allows it to function as a 'semantic VAD' (Voice Activity Detector), determining not just when sound stops, but when a user's semantic thought is complete, enabling more natural turn-taking.

To ensure rigorous testing, the team also released SoulX-Duplug-Eval, an extended evaluation benchmark that improves upon existing tests with better bilingual coverage. Experimental results confirm that systems built with SoulX-Duplug achieve lower latency in streaming dialogue and outperform current full-duplex models in both turn management and speed. By open-sourcing both the module and the evaluation suite, the researchers aim to accelerate development in the field, moving AI assistants closer to fluid, interruption-friendly conversations that mimic human interaction. The paper has been submitted for review at Interspeech 2026.