Tencent Hunyuan open-sources Hy-SimulMT, a speech interpreter with about 3 seconds of lag

Tencent Hunyuan has released and open-sourced Hy-SimulMT, a simultaneous interpretation model that turns Chinese speech into English speech, and English into Chinese, while the speaker is still talking. It clones the speaker's voice in real time, so listeners hear the translation in something close to the original voice. The model is live first in the Hy Translate app, where it is free to use, according to IT Home and the project's GitHub and Hugging Face pages.

The number that matters is latency. Tencent measures it with StreamLAAL, a standard metric for streaming translation lag, and reports about three seconds on the long-form RealSI benchmark, the lowest among the systems it compared. The project notes that StreamLAAL is not the same as the time until the first translated speech is heard. For comparison, IT Home puts the typical lag of professional human interpreters at four to five seconds.

On quality, Hy-SimulMT's GEMBA score on long-form speech beat the open models in the comparison and four of the five commercial APIs tested. On short-form sets it scored 90.3 on FLEURS, 75.5 on CoVoST2, 81.2 on ACL6060 and 84.7 on RealSI-Sentence, the last being the highest of every listed system.

The team also compared the model with human interpreters on 18 recordings of about six minutes each, captured from September 2026 onward across 10 domains. Six interpreters took part, three holding China's top CATTI Level 1 interpreting certificate and three holding Level 2, and the best result from each level was used. Tencent calls this a small-scale assessment.

Wang Weiwei, vice dean of the School of Interpreting and Translation Studies at Guangdong Foreign Studies University, said in remarks published with the release that the system speaks first and corrects itself afterward, much like a human interpreter, and that AI interpreting is moving from helping people understand toward usable communication in low-risk settings.

The released model is a 30B-A3B mixture-of-experts system, with FP8 and GGUF builds and an Interleave variant that also outputs the source transcript. The Hugging Face card lists an Apache 2.0 license. Tencent ships two tools with it: SimulInfer, a serving stack built on vLLM with a web interface, a WebSocket API and optional text-to-speech, and SimulTrain for further training on domain data. Support in llama.cpp has not yet been merged.

For the field, the signal is less about benchmark scores than about latency: real-time speech translation has moved to within a second or two of what professional human interpreters achieve, and the model, serving stack and training tooling shipped together. Three-second interpretation plus voice cloning makes talking through a translator feel closer to a natural conversation, though the same design means a mistranslation arrives in the speaker's own voice, with all the labeling and accountability questions that follow.

What to watch next: when backends such as llama.cpp pick the model up, how latency and stability hold under real concurrent load, and whether the small human evaluation reproduces across more language pairs and domains. Tencent frames the target as low-risk settings, and medical, legal or diplomatic use would clearly need far stricter validation.