On November 13, Tencent Hunyuan released Hy Realtime, a real-time voice model that handles the listening and speaking in voice calls. This generation adds two capabilities: calling tools in the middle of a conversation (for now, mainly deciding whether to search the web), and taking spoken song requests and singing the result back.
The model is live in Yuanbao voice calls and in Tencent customer service. Yuanbao users can update to the latest version and pick the "Chuyu" voice in a voice call to try it, and Tencent has also put up a web demo that exposes part of the feature set. The developer API will arrive through Tencent Cloud; the official wording is "soon."
Understanding half a sentence, and searching the web
The Hunyuan team put most of this release's effort into comprehension. In real calls, users often mumble, trail off mid-sentence or change their mind halfway, and the model has to use earlier context to fill in the intent. Tencent says accuracy on this kind of incomplete input has improved "significantly" and that context holds up across long multi-turn conversations, but it gives no figures.
On tool use, the announcement is fairly restrained: the model judges from context whether a web search is needed, and handles the search decision, the tool call and the answer in a single flow. In the reinforcement learning stage, the team widened the reward target from "is the final answer right" to the whole calling process, and applies stepped penalties to unnecessary retrievals. The aim is to stop the model from running off to search every time a user casually asks about the weather or a score.
The singing feature was built with the TME Lingyin AI team at Tencent Music. Users can combine singer, album, genre, language, release year and popularity when requesting a song, for example "an English folk song from between 2010 and 2020," and the model matches a track and sings it directly. As for quality, Tencent says only that it showed a "significant advantage" in internal expert A/B tests; the evaluation details have not been published.
The numbers in the announcement that can be checked
The technical section reveals little quantitative information, but three points are fairly specific:
- Pre-training used speech data on the scale of tens of millions of hours, with tasks widened from plain speech recognition to audio understanding, music understanding and speech translation;
- It comes with 51 text-only evaluation sets and more than 50 speech evaluation sets, used to watch whether "text IQ" drops once it is converted to speech;
- In the supervised fine-tuning stage, the data pipeline for rewriting text into spoken style was rebuilt, and the effective retention rate of high-quality text data transferred to speech tasks rose from 21% to 69%.
That last figure targets a familiar weakness of end-to-end voice models: on the same base model, a question that gets answered correctly in a typed chat gets a shallower answer by voice. Tencent's approach is to have both the Hunyuan text model and a voice agent model act as teachers, jointly guiding Hy Realtime at a controlled ratio. On the speech generation side, it uses a streaming design that synthesizes speech while the text is still being produced, intended to cut first-response latency. The announcement does not say how many milliseconds that latency is.
What it means for users in China
Most of the AI voice-call entry points Chinese users rely on sit inside the assistant apps of big tech companies. Hy Realtime landing first in Yuanbao and Tencent customer service makes the path clear: refine the model on real call volume from Tencent's own products, then sell it through Tencent Cloud to companies building customer service, coaching, in-car and smart-hardware products.
For ordinary users, the most visible change is that Yuanbao voice calls now offer song requests, and the assistant will also search the web more proactively when looking things up. For enterprise developers, the things to watch come later: how the API will be priced, whether custom voices are supported, whether singing will reach the API, and how song copyright is handled. The announcement mentions none of these, and they can only be confirmed once the Tencent Cloud product page goes live.
The Hunyuan team lists full-duplex interaction, long-term personalized memory and foreground-background agent collaboration as next steps. Full-duplex means the user and the model can speak at the same time and interrupt each other at any moment; real-time voice products on the market are all still refining this, and Tencent does not claim to have achieved it this time either.
Sources: Tencent Hunyuan official announcement, Yuanbao app voice call feature documentation, CocoLoop, Tencent Cloud web demo; the 21% and 69% retention figures and the evaluation set counts follow the Tencent Hunyuan announcement.