Tencent Hunyuan open-sources Hy-SimulMT for real-time Chinese-English speech translation
Article timeline
The articles in this story. Times are ET.
ModelScopeOfficialModelScope releases open Hy-SimulMT for live Chinese-English speech translationAIModelScope announces Hy-SimulMT, an open model for simultaneous Chinese-English translation of live speech, with text and speech output. The post says it has the lowest StreamLAAL among compared systems in the RealSI long-form evaluation and beats four of five commercial APIs on GEMBA quality. It is Apache-2.0 licensed, offers FP8 and GGUF variants, and needs a configured TTS service for speech output and voice cloning.
SGLangOfficialSGLang Omni adds Tencent's Hy-SimulMT for real-time speech translationAISGLang says Tencent's Hy-SimulMT speech translation model is coming to SGLang Omni. It supports streaming Chinese-English audio input with translated text output and runs BF16 inference on a single H200. The post gives the serve command sgl-omni serve --model-path tencent/Hy-SimulMT-30B-A3B --enable-realtime.
- PandailyNewsPickTencent Hunyuan open-sources Hy-SimulMT, a live speech interpreter with about 3 seconds of lag
AITencent Hunyuan has open-sourced Hy-SimulMT, a model that translates Chinese and English speech into the other language in real time while cloning the speaker's voice. Tencent reports about 3 seconds of latency on the RealSI benchmark using the StreamLAAL metric, which the project notes is not the same as time to first translated speech. The 30B-A3B mixture-of-experts model is available under an Apache 2.0 license, and it is live first in the Hy Translate app.
Tencent HyOfficialTencent open-sources Hy-SimulMT for real-time Chinese-English interpretingAITencent Hunyuan releases Hy-SimulMT, an open-source simultaneous interpreting model for Chinese and English, on Hugging Face, ModelScope and GitHub. The company says it runs with about 3-second latency, supports real-time voice cloning, and outperforms most open-source models and commercial APIs.
