Tencent Hunyuan Open-Sources Hy-SimulMT, a Simultaneous Interpretation Model with ~3-Second Latency

On October 26, Tencent Hunyuan released and open-sourced Hy-SimulMT, a simultaneous interpretation model supporting real-time speech-to-speech translation between Chinese and English. The model processes incoming audio as it is spoken, generating translations and speech output on the fly, and supports real-time voice cloning to reproduce the speaker's voice and tone.

According to official measurements, Hy-SimulMT responds in about 3 seconds, compared with an average of 4 to 5 seconds for professional human interpreters. In quality tests, the model outperformed mainstream open-source simultaneous interpretation models and most commercial simultaneous interpretation APIs. On RealSI, a commonly used long-form audio open-source evaluation set, Hy-SimulMT achieved the best translation quality among open-source models while showing the lowest latency.

To assess real-world performance, the team collected Chinese and English audio from the internet, covering ten domains including software AI, medicine, manufacturing and energy, natural science, agriculture and food, business management, law, education, culture and tourism, and sports and lifestyle, across scenarios such as speeches, interviews, and meetings. Six professional interpreters holding CATTI Level I or Level II interpreting certificates were invited for comparison. Hy-SimulMT matched or exceeded the human interpreters in translation quality while delivering faster responses.

Hy-SimulMT is now available for free in the Tencent Hy translation app.