Tencent's Voice Model Decides When to Search and Can Sing
Tencent's Hunyuan team has released Hy Realtime, a voice model that breaks into song on request, judges for itself when a spoken question needs a web search, and is already answering calls inside the company's Yuanbao assistant.
The model is live in voice calls in Yuanbao and in Tencent's customer service, and the company says it will be offered through an API on Tencent Cloud, without giving a date or a price for that access.
And it is built for the way people actually speak, according to Tencent, reconstructing intent from a sentence that stops halfway or corrects itself, with better tracking of context across a long conversation.
Tool use sits in the same loop, so the model decides whether a search is needed, runs the search and speaks the answer as one process rather than handing the request to a separate system.
The singing was built with an AI team at Tencent Music Entertainment. A user can ask for songs by artist, album, genre, language, release year or popularity, stack those conditions and refine the request in later turns, and Tencent's example has a user ask for an English folk song released between 2010 and 2020. The company says it tuned the singing for expressiveness, natural phrasing and breath, and for melody and rhythm.
But the evidence for the model's quality is Tencent's own. It reports gains in internal blind tests, in an internal reasoning set and on several public voice benchmarks, describing its results on those public tests as leading, without giving sample sizes or test settings.
Output quality is where the company claims most, with improvements in audio, pronunciation, prosody and emotional expression, and more natural shifts in tone over the course of a conversation.
The model can be tried in the latest Yuanbao app by choosing the voice named Chuyu in a call, or through a demo on Tencent Cloud, and the team says duplex interaction, long term memory and coordination between a foreground agent and background agents come next.
So Tencent has moved the search decision and the singing inside the conversation, and the benchmark claims stay the company's own until outside tests appear.
Hy Realtime is live in Yuanbao voice calls and Tencent customer service
It decides when a question needs a web search and runs the search itself
Singing was built with Tencent Music Entertainment's AI team
Tencent claims leading public benchmark results without sample sizes
An API on Tencent Cloud is promised next
Source: Pandaily
