On the TTS Arena leaderboard hosted on Hugging Face, often dubbed the "Olympics" of AI, the Chinese AI voice model"Wusheng Vocu V3" rose to the top of the global rankings after blind voting by tens of thousands of users worldwide, outperforming well-known rivals such as US-based Inworld AI. Behind this achievement are Xie Weiduo, an entrepreneur born in the 2000s, and Guangzhou Shuogu Technology Co., Ltd., the company he leads.

According to Xie, "Wusheng Vocu" is capable of deeply understanding the emotions in a text and accurately rendering crying, laughing, singing and other expressive vocal effects. With just a three-second voice sample, it can achieve instant voice cloning with a similarity rate of more than 95%. Within only three months of its launch, Wusheng has attracted more than one million users and recorded tens of millions of visits. It has also completed the filing process for deep synthesis service algorithms.

At the beginning, Xie just wanted to find a pleasant-sounding voice for his AI virtual streamer, Moe. Disappointed with all the voice solutions available on the market, he decided to build one himself. He threw himself into learning about speech synthesis and experimented with using a generative AI-based architecture for speech synthesis, encoding voices into a text-like form before reconstructing them through algorithms. In 2023, when Moe spoke with a new voice, it unexpectedly drew an enthusiastic response from online users. Xie later suspended his studies and returned to Guangzhou to start his business, founding Guangzhou Shuogu Technology Co., Ltd. Today, his AI voice model has helped drive the company's valuation to hundreds of millions of yuan. The company is now moving steadily ahead with fundraising and has begun to generate profits.

这位广东“00后”将AI语音模型做到全球第一

在被誉为AI“奥林匹克”的HuggingFace的TTS Arena榜单上，中国AI语音模型“悟声Vocu V3”在数万名全球用户的盲测投票中登顶世界第一，将美国Inworld等知名同类产品甩在身后。创造这一奇迹的，是“00后”谢伟铎和他带领的广州烁谷科技公司。

谢伟铎介绍，“悟声Vocu”能深度理解文本情感，精准演绎哭、笑、歌唱等。仅需3秒的声音样本，就能实现相似度超95%的瞬时克隆。上线短短3个月，“悟声”用户量突破百万，访问量达千万级，并顺利通过了深度合成服务算法备案等。

在最开始，他只是想给自己的AI虚拟主播“木几萌”找一个好听的声音。在对市面上所有的语音方案失望后，谢伟铎打算“自己做一个”。他恶补语音合成知识，尝试用生成式AI架构做语音合成，将声音编码成类似文本的形式，再通过算法还原。2023年，“木几萌”搭载着新声音开口时，意外获得了网友们的热烈反响。后来谢伟铎选择休学回广州创业，成立“烁谷科技”。如今，他的AI语音模型已获数亿元估值，正稳步推进融资事宜，并走上了盈利之路。

文 | 记者 黎秋玲
图 | 记者 刘志勇
译｜郑奕玲
英文审校｜林佳岱