Tencent unveils open-source simultaneous translation model with 3-second response time

Tencent has released and open-sourced its Hy-SimulMT simultaneous translation model, which delivers real-time voice-to-voice translation between Chinese and English with a response time of just 3 seconds—faster than human interpreters. The model, launched in its Hy Translation App, supports real-time audio input, translation, and voice output, along with voice cloning to replicate the speaker’s tone and pitch. According to Tencent, Hy-SimulMT outperforms most open-source and commercial translation APIs in both speed and accuracy. In tests on the RealSI dataset, the model achieved the lowest delay while maintaining high translation quality. Human experts, including three CATTI Level 1 and three CATTI Level 2 certified interpreters, evaluated the model, finding its translation quality matched or exceeded human performance, with significantly lower latency.