Tencent Hunyuan releases Hy Realtime voice model with tool calling and singing, live in Yuanbao

Tencent Hunyuan has released Hy Realtime, a real-time voice model that can call tools and sing during spoken conversation. The team announced it on November 13. The model is already live in voice calls on Tencent Yuanbao and in Tencent's customer service. Tencent says it will soon be offered through an API on Tencent Cloud. Pricing is not yet disclosed.

Tool calling inside the voice loop

Hy Realtime decides, from the conversation and the user's intent, whether a web search is needed. According to Tencent, the search decision, the tool call and the spoken answer are handled as one process rather than as separate steps.

Handling messy speech

Tencent says the model is built for how people actually talk. Users may mumble, stop halfway through a sentence or correct themselves mid-thought. The model uses the surrounding conversation to reconstruct what they meant. Tencent also says multi-turn instruction following and context management have improved, so the model keeps track of intent through longer conversations.

On output, Tencent says the model improves on its predecessor in audio quality, pronunciation, prosody and emotional expression, with more natural shifts in tone over a conversation.

Singing, built with Tencent Music Entertainment

The singing capability was developed together with an AI team at Tencent Music Entertainment. Users can request songs by:

• artist
• album
• genre
• language
• release year
• popularity

Conditions can be combined, and requests can be refined in follow-up turns. In Tencent's example, a user asks for an English folk song released between 2010 and 2020, and the model picks a match and sings it. Tencent says it tuned the singing for expressiveness, natural phrasing and breath, and for melody and rhythm.

Benchmarks

Tencent reports gains in internal expert blind tests of search-and-answer tasks, in an internal reasoning test set for non-thinking mode and on several public voice benchmarks. It describes its overall results on those public benchmarks as leading. These are company evaluations. No numeric scores, sample sizes or test settings have been published.

Availability

• Yuanbao: choose the voice named Chuyu in voice calls in the latest Yuanbao app.
• Web demo: available on Tencent Cloud.
• API: coming soon on Tencent Cloud. Pricing and context window are not yet disclosed.

Tencent lists three next steps: duplex interaction, long-term personalized memory, and coordination between a foreground conversational agent and background agents.

What this means

Hy Realtime moves search decisions into the voice model itself. Tencent says the model decides whether to search, makes the tool call and speaks the answer as a single process. For builders, that matters most once the Tencent Cloud API ships, since it would show whether tool use can be handled inside one voice model. Until then, the only way to evaluate the model is through Yuanbao and the web demo.

The benchmark claims cannot yet be independently weighed. Tencent gives no numbers for its internal blind tests or public benchmark results, and the term "leading" is its own description. The singing feature, developed with an AI team at Tencent Music Entertainment, is the most distinctive part of the release. Tencent's stated roadmap of duplex interaction and long-term memory suggests the current version is still a half-duplex, per-session system.

Source: Pandaily, citing IT Home's report of the official announcement and the Tencent Cloud web demo.