Tencent Hunyuan open-sources Hy-SimulMT speech-to-speech interpreter with ~3s lag

Tencent Hunyuan has released and open-sourced Hy-SimulMT, a simultaneous interpretation model that converts Chinese speech to English and vice versa while cloning the speaker's voice in real time. It reports about 3 seconds of latency on the RealSI benchmark by StreamLAAL, versus 4–5 seconds for professional human interpreters. The 30B-A3B mixture-of-experts model is free in the Hy Translate app under an Apache 2.0 license.

• About 3s latency by StreamLAAL on RealSI vs 4–5s for human interpreters
• 30B-A3B MoE model with FP8 and GGUF builds under Apache 2.0
• Short-form scores: FLEURS 90.3, CoVoST2 75.5, RealSI-Sentence 84.7
• Ships with SimulInfer on vLLM and SimulTrain for domain fine-tuning