Tencent Hunyuan launches real-time voice model Hy Realtime: can speak and sing, and even understand last-minute changes

Beating AI News Flash: Tencent Hunyuan launches its new real-time voice model Hy Realtime, adding singing and web search capabilities during voice conversations, and has integrated it into Yuanbao voice calls and Tencent customer service.

Users can directly use voice to have the AI search for the latest information, and can also request songs by artist, genre, language, and release year. For example, asking it to sing an English folk song from between 2010 and 2020, the model can select a song based on context and sing it. The singing feature was jointly developed by Hunyuan and Tencent Music's TME Lingyin AI team.

The new model also focuses on solving the problem of voice AI easily "losing intelligence when it starts speaking." Knowledge and reasoning capabilities originally held by large text models may suffer losses after being converted into voice conversation models. Tencent redesigned the training data production process, claiming that the effective transfer retention rate of high-quality text data to voice tasks increased from 21% to 69%, about 3.3 times the previous level.

Hy Realtime also strengthens its understanding of real spoken language. When users speak vaguely, say only half a sentence, or change their wording midway, the model can restore intent based on context. When generating responses, it can organize text and synthesize voice at the same time, reducing waiting time.

Currently, users can experience it by selecting the "Chuyu" voice in voice calls in the latest version of the Yuanbao App. The Tencent Cloud API is planned to open soon, while duplex interaction and long-term personalized memory remain directions for subsequent upgrades.