Tencent releases open-source AI model for real-time translation

Tencent has released the source code and model weights for Hy-SimulMT, an AI system designed for simultaneous interpretation. The tool processes incoming speech streams to provide real-time translation in text or audio format, utilizing voice cloning to maintain the speaker's original vocal characteristics.

Performance metrics provided by Tencent indicate a streaming translation lag of approximately three seconds on the RealSI audio benchmark. The company claims these results outperform several competing open-source models and the majority of comparable commercial APIs, though full speech generation functionality requires additional text-to-speech integration.

The release includes the necessary training tools and inference software for developers. The technology is currently being demonstrated in the company’s Hy Translate application, marking a shift toward broader external access for Tencent’s language processing tools.