正文 · AI 翻译
Thanks to IT Home readers Qingchuan Wenhe and Huanan Wu Yanzu for the tips!
IT Home reported on October 26 that Tencent today released and open-sourced Hy-SimulMT, a Hunyuan simultaneous interpretation large model characterized by low latency and high accuracy. It supports real-time speech-to-speech translation between Chinese and English: a speaker's audio can be fed into Hy-SimulMT in real time, and the model translates as it listens, producing translated text and speech in real time. It also supports real-time voice cloning, closely reproducing the speaker's voice and tone. The model debuted in Tencent's Hy Translation app, where the service is available to try for free.
IT Home learned from the official introduction that low latency is a major feature of the model. In actual tests, Hy-SimulMT's response time was around 3 seconds, compared with an average of 4-5 seconds for professional human interpreters. Evaluation results showed that its translation performance surpassed that of mainstream open-source simultaneous interpretation models on the market and even most commercial simultaneous interpretation APIs.
According to the official account, Hy-SimulMT demonstrated good simultaneous interpretation performance on RealSI, an open-source long-audio evaluation dataset commonly used in the industry, as shown below:
StreamLAAL on the horizontal axis is a simultaneous interpretation latency metric. It reflects, in seconds, the delay between the input audio and the model's translated output; lower is better. The vertical axis represents translation quality, scored using GPT6.1-Sol; higher is better. Hy-SimulMT achieves the best performance among open-source models while maintaining the lowest latency, and even outperforms most commercial APIs.
Hy-SimulMT also performed well on a short-audio test set, as shown below:
To assess Hy-SimulMT's actual capabilities more accurately, the team collected a broad range of newly posted Chinese and English audio from the internet, within the bounds of compliance requirements. The material covered 10 fields—software and AI, medicine, manufacturing and energy, natural sciences, agriculture and food, business management, law, education, culture and tourism, and sports and everyday life—and included speeches, interviews, meetings and other settings.
Using this dataset, the team tested Hy-SimulMT's performance and also asked several human simultaneous interpretation experts to interpret the audio.
There were 6 human simultaneous interpretation experts in total: 3 held CATTI Level 1 interpreting certificates, and 3 held CATTI Level 2 interpreting certificates. All had many years of practical simultaneous interpretation experience. The evaluation compared Hy-SimulMT's translations with those of the human experts on latency and translation quality; for the human experts, it used the highest score among 3 people. In this capability assessment, Hy-SimulMT reached the level of human simultaneous interpretation experts in translation quality, with a higher score, and was faster, with lower latency.
References
HuggingFace
Modelscope
Github
Advertising notice: External links in this article, including but not limited to hyperlinks, QR codes and passcodes, are provided to convey more information and save time in selecting information. The results are for reference only. All IT Home articles containing external links include this notice.