Tencent Open-Sources Simultaneous Interpretation Large Model: 3.3-Second Latency, 1.2 Seconds Faster than CATTI Level 1 Experts

Tencent Hunyuan has released and open-sourced Hy-SimulMT, a simultaneous interpretation large model. It takes Chinese and English speech input and directly outputs speech. The model interprets as it listens, synchronously outputting translation and voice, while also cloning the speaker's voice in real time, reproducing both the speaker's voice and tone. It has made its debut in Tencent's Hy Translate app, where it is free to try.

Latency is the key metric for simultaneous interpretation. Hy-SimulMT has a measured response time of about 3 seconds, while human professional interpreters average 4–5 seconds. Tencent compared it with six certified simultaneous interpretation experts: three with CATTI Level 1 and three with CATTI Level 2. CATTI is the China Accreditation Test for Translators and Interpreters. The tests covered ten fields, including software/AI, medicine, manufacturing/energy, and business management. Hy-SimulMT's StreamLAAL latency was 3.3 seconds, compared to 4.5 seconds for CATTI Level 1 experts and 4.7 seconds for Level 2 experts. Its translation quality matches that of human simultaneous interpretation experts.

On the long-form audio evaluation set RealSI, the model scored about 86 points with a latency of approximately 3.1 seconds. Its latency was the lowest among compared models, and its quality was the best among open-source models, surpassing most commercial simultaneous interpretation APIs. Latency is measured by StreamLAAL, and translation quality is evaluated with GEMBA, with GPT-6.1-Sol as the scoring model.

In the four short-form audio tests, Hy-SimulMT scored 84.7 on RealSI-Sentence, the highest overall. It scored 90.3 on FLEURS and 81.2 on ACL6060, ranking first among open-source models. Gemini 3.5 LiveTranslate took first place on FLEURS and ACL6060 with 92.4 and 82.6, respectively.

The strategy for controlling latency is called “speak first, repair later.” The model starts translating based on the information it has received so far, then revises its output as more input arrives. The result sounds close to the self-corrections made by human interpreters.

Wang Weiwei, Vice Dean of the School of Interpreting and Translation Studies at Guangdong University of Foreign Studies, and Professor Zhong Weihe, former Executive Vice President of the Translators Association of China, commented: Hy-SimulMT performs impressively in speech recognition stability, simultaneous interpretation response speed, and target-language delivery quality, helping move machine simultaneous interpretation into real-world scenarios.

The code is now available on HuggingFace, ModelScope, and GitHub. Developers can run it themselves, and ordinary users can open Tencent's Hy Translate app to use it. The barrier to simultaneous interpretation is being torn down.