Tencent Releases Hy-SimulMT-30B-A3B Real-Time Speech-to-Speech Translation Model
Tencent Hunyuan released Hy-SimulMT-30B-A3B, a 30-billion parameter real-time simultaneous speech-to-speech translation model that supports bidirectional communication between Chinese and English. It continuously processes incoming audio streams to generate translated text and speech while preserving the original speaker's vocal characteristics through voice cloning. This release provides the open-source community with an advanced tool for live communication settings like meetings and speeches, achieving lower latency and higher translation quality than many commercial APIs. It bridges language barriers while maintaining personal vocal identity during cross-lingual conversations. In evaluation, the model achieved a StreamLAAL of approximately 3 seconds, representing the lowest latency among compared systems, and outperformed several commercial APIs on GEMBA scores. The repository provides the SimulInfer inference service and SimulTrain training tools, though speech output and voice cloning require a separately configured TTS service.
## BACKGROUND
Simultaneous speech-to-speech translation (Simul-S2ST) is a technology that outputs target speech while receiving streaming speech inputs, which is critical for achieving low-latency real-time communication. StreamLAAL measures the latency of simultaneous translation systems, helping researchers evaluate the speed and responsiveness of real-time speech processing models.