Tencent Hunyuan open-sources AI simultaneous interpretation model: translates while listening with about 3-second latency, and can also clone the speaker's voice.

Beating AI News Flash: Tencent Hunyuan has open-sourced its simultaneous interpretation model Hy-SimulMT, supporting bidirectional real-time translation between Chinese and English. Without waiting for the speaker to finish a sentence, the model can generate translations while listening.

After integrating an additional TTS model, it can also clone the original speaker's voice and broadcast in another language. The official deployment solution is based on Qwen3-TTS.

In official RealSI long-audio tests, it achieved a streaming translation latency of approximately 3 seconds, the lowest among all tested systems. Translation quality surpassed all tested open-source models and 4 of 5 commercial APIs. The 3 seconds here does not include the full speech synthesis and broadcasting time.

The team also recruited 6 certified professional interpreters to conduct side-by-side tests on 18 recordings of real speeches, interviews, and more. In this small-scale test, the model's translation quality reached professional simultaneous interpretation standards, with typically lower latency.

The model has approximately 30 billion total parameters, with about 3 billion activated per inference. Weights, inference, and training tools are all open-sourced, and the Tencent Hy Translation App also offers free trials.