What changed | Translation can begin before the speaker finishes

On October 26, 2026, Tencent Hunyuan released Hy-SimulMT, a real-time model for Chinese-to-English and English-to-Chinese interpretation. It takes incoming audio and produces a translation as the speaker continues. The project includes the SimulInfer inference service and training tools. You can watch the team's examples to hear how the system performs. Spoken output requires a separately configured text-to-speech (TTS) service. Project and release notes

Why it matters | Listen to what happens when a speaker corrects a time

Imagine an event host saying, “We leave at two this afternoon … actually, make that two thirty.” Checking only the final wording misses something the audience experiences live. How far had the speaker got when “two” first appeared? When the speaker corrected the time, did the subsequent translation make it clear that the departure was at 2:30?

For a live-caption trial, record those two observations separately. Listen to the original, mark when the first translation appears, then check the action and time conveyed by the end. A system might start translating quickly and miss the correction. Another might get the time right but leave the audience waiting at the outset. The project publishes its own evaluations and demonstrations. The departure announcement is a sample you can try yourself.

How to try it | Start with a short recording

  1. Watch a published example. In the official repository, choose a clip in a language you understand. Compare the speaker's delivery with the timing of the translation. The repository also links to the Hy Translate mobile app for Android and iOS. Examples and app links
  2. Try your own material. If you already have the inference service running, wait until it reports that it is ready. On the server machine, open http://127.0.0.1:8080/, choose Chinese to English or English to Chinese, then upload a short recording or use the microphone. Choose Stop and finish when you are done. Remote microphone access requires HTTPS. To set up the documented vLLM route yourself, first check that you have Linux, Python 3.12 or later, Git, an NVIDIA GPU with enough memory and a compatible CUDA driver. Then follow the repository's deployment instructions. Requirements and web interface
  3. Check the output you need. For captions, check whether “two thirty” came through and how long the listener waited. For interpretation through headphones, configure speech synthesis as well and listen to the spoken result. You can do the caption check with text alone.