Tencent opens Hy-SimulMT weights for translation while speakers talk
Tencent's Hunyuan team has released Hy-SimulMT weights for streaming Chinese-English speech translation. The model produces translated text before a speaker finishes, while spoken output and voice cloning require a separate text-to-speech service. Supporting inference and training tools accompany the release; published performance comparisons remain Tencent's evaluations rather than independent measurements.
Artificial Intelligence··Morning
Translation text arrives while speech continues
Tencent's Hunyuan AI team released Hy-SimulMT weights for simultaneous translation between Chinese and English. The model accepts incoming speech and produces translated text while the speaker continues. A base version returns the translation, while Interleave returns recognized source text alongside it. The downloadable weights are offered under Apache-2.0 licensing. The release targets translation that can proceed before an entire sentence is complete.[1], [2]
Spoken output needs a separate synthesis service
The checkpoint itself outputs text. Producing spoken translations or imitating a speaker's voice requires a separately configured text-to-speech service. SimulInfer, the accompanying inference service and web interface, accepts uploaded audio or microphone input. SimulTrain supplies training tools. Technical reporting describes a mixture-of-experts design, in which selected parts of the model participate in processing each input, rather than every parameter being active throughout.[1], [2]
Local runtime support remains incomplete
The GGUF model-file builds still awaited support in the main llama.cpp project, the local inference software. A broader language list in repository metadata does not establish speech support beyond Chinese and English. Tencent's published comparisons use translation tests and a small collection of recordings interpreted by professionals. Those results remain vendor evaluations, without independent measurements of the released model's translation performance.[1], [2]
