Tencent's Hunyuan open source simultaneous interpretation large model Hy-SimulMT: The delay is lower than that of human interpreters, and the quality is comparable to professional interpretation. On October 26, Tencent's Hunyuan simultaneous interpretation large model Hy-SimulMT was released and open source, supporting Chinese and English speech-to-speech real-time translation: the speaker's audio can be input into the model in real time, and the translation and speech can be output simultaneous

Tencent's Hunyuan open source simultaneous interpretation large model Hy-SimulMT: lower latency than human translators, quality comparable to professional interpretation On October 26, Tencent's Hunyuan simultaneous interpretation large model Hy-SimulMT was released and open sourced, supporting real-time Chinese and English speech-to-speech translation: the speaker's audio can be input into the model in real time, and the translation and speech are output simultaneously while listening and interpreting. It also supports real-time timbre cloning to restore the speaker's voice and tone. The model has been launched for free on Tencent Hy Translation App, and the open source warehouse has been launched simultaneously with HuggingFace, ModelScope and GitHub. Low latency is the core indicator of this model. The measured response is about 3 seconds, which is lower than the average level of 4-5 seconds for human professional translators. In a comparison with six certified simultaneous interpretation experts (3 CATTI Level 1 and 3 CATTI Level 2) on audio in ten fields including software AI, medicine, manufacturing energy, and business management, Hy-SimulMT's StreamLAAL delayed 3.3 seconds, which was better than the 4.5 seconds of CATTI Level 1 experts and 4.7 seconds of Level 2 experts. The translation quality reached the level of human simultaneous interpretation experts. On the long audio evaluation set RealSI, StreamLAAL measures latency and GEMBA measures translation quality (scored by GPT6.1-Sol). Hy-SimulMT scored about 86 points with a latency of about 3.1 seconds. It has the lowest latency and the best quality among open source models, surpassing most commercial simultaneous interpretation APIs. Among the four short audio tests, its RealSI-Sentence score of 84.7 was the highest in the field, FLEURS90.3 and ACL606081.2 ranked first among open source models, and Gemini3.5-LiveTranslate ranked first in individual categories with FLEURS92.4 and ACL606082.6. The model uses the strategy of "speaking first, patching later" to control delay, and the patching listening experience is close to the translator's self-correction. Wang Weiwei, deputy dean of the College of Translation and Interpretation at Guangzhou Foreign Studies University, and Professor Zhong Weihe, former executive vice president of the Translators Association of China, believe that its speech recognition stability, simultaneous interpretation response speed and target language broadcast quality are all very good, pushing machine simultaneous interpretation into real scenes. via AI News (author: AI Base)