IT之家 10 月 26 日消息,腾讯混元同声传译大模型 Hy-SimulMT 今日发布并开源,具备低时延、高准确率的特征,兼容中英语言之间语音到语音实时翻译 —— 说话人的音频可以实时输入到 Hy-SimulMT 模型中,模型会边听边译,实时给出译文和语音,并兼容实时音色克隆,可以高度还原说话人声音和语气。该模型在腾讯 Hy 翻译 App 首发,免费给出服务感受。
IT之家从官方介绍获悉,低时延是模型的一大亮点,混元同声传译大模型 Hy-SimulMT 的响应时间实测在 3 秒左右,人类专业译员平均水平在 4-5 秒。测评结果表明,该模型的翻译成效超过了市面上主流的开源同传模型,甚至超过大部分的商用同传 API。
官方称,在业界常用的长音频开源实测集 RealSI 上,Hy-SimulMT 展示出了较好的同传翻译成效,见下图:
横轴里的 StreamLAAL 表示同传翻译的延迟指标,反映的模型输出翻译结果相比输入音频的延迟秒数,越低越好。纵轴表示翻译质量,使用 GPT6.1-Sol 开展打分,越高越好。Hy-SimulMT 能做到延迟最低的情况下,成效达到开源模型最佳,甚至超过大多数商业 API。
在短音频测试集上,Hy-SimulMT 也有较好表现,见下图:
为了更准确评估 Hy-SimulMT 的真实水平,该团队在合规允许范围内,广泛收集了互联网上新浮现的中文和英文音频,包含软件 AI、医学、制造能源、自然科学、农业食品、商务管理、法律、教育、文化文旅、体育生活共十个领域,涉及演讲、访谈、会议等多个情形。
在这份数据上,该团队不仅测试了 Hy-SimulMT 的成效,也请多位人类同传专家对这些音频开展同声传译。
人类同传专家一共 6 位,其中 3 位具备 CATTI 一级口译证书,3 位具备 CATTI 二级口译证书,所有专家都有多年的同传实战经验。测评对 Hy-SimulMT 和人类同传专家的翻译内容在延迟和翻译质量上比较,其中人类同传专家选取的是 3 个人中的最高分,在这次能力定位测评中,Hy-SimulMT 在翻译的质量达到人类同传专家水平(得分更高),速度更快(时延更低)。
参考
本文来源:https://www.ithome.com/1/018/168.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Thanks to IT Home users Qingchuan Wenhe and South China Wu Yanzu for the tip!
IT Home, October 26 – Tencent Hunyuan’s simultaneous interpretation large model Hy-SimulMT was released and open-sourced today, featuring low latency and high accuracy, and supporting real-time speech-to-speech translation between Chinese and English—speaker audio can be fed into Hy-SimulMT in real time; the model listens and translates simultaneously, delivering both translated text and voice output on the fly, and it is compatible with real-time voice cloning, faithfully reproducing the speaker’s voice and tone. The model debuts in the Tencent Hy Translation App and is available for free trial.
According to the official briefing, low latency stands out as a major highlight. The measured response time of Hunyuan’s Hy-SimulMT is around 3 seconds, whereas professional human interpreters average 4–5 seconds. Benchmarks show the model’s translation performance surpasses mainstream open-source simultaneous-interpretation models and even most commercial simultaneous-interpretation APIs.
The company stated that, on the widely used RealSI open-source long-audio benchmark, Hy-SimulMT demonstrated strong simultaneous-interpretation performance, as shown below:
On the horizontal axis, StreamLAAL denotes the latency metric for simultaneous interpretation, reflecting the seconds of delay between the model’s translation output and the input audio—lower is better. The vertical axis represents translation quality, scored using GPT6.1-Sol—higher is better. Hy-SimulMT achieves the lowest latency while delivering the best performance among open-source models and outperforms the majority of commercial APIs.
On the short-audio test set, Hy-SimulMT also performs well, as shown below:
To more accurately evaluate Hy-SimulMT’s real capabilities, the team—within compliant limits—collected a broad range of newly surfaced Chinese and English audio from the Internet, spanning ten domains: software AI, medicine, manufacturing and energy, natural sciences, agriculture and food, business management, law, education, culture and tourism, and sports and lifestyle, covering various scenarios such as speeches, interviews, and conferences.
On this dataset, the team not only tested Hy-SimulMT’s performance but also invited multiple human simultaneous-interpretation experts to interpret the same audio clips.
There were six human simultaneous interpreters in total, three holding CATTI Level 1 Interpretation Certificates and three holding CATTI Level 2 Interpretation Certificates; all have many years of real-world simultaneous-interpretation experience. The assessment compared Hy-SimulMT and the human experts’ outputs in terms of latency and translation quality, taking the highest score among the three selected human interpreters. In this capability benchmark, Hy-SimulMT achieved human-expert-level translation quality (scoring higher) with faster speed (lower latency).
Reference


