It is past midnight in Shenzhen, and a customer is on the phone with Tencent support, explaining a billing problem in a tired, rambling voice. On the other end there is no sleepy human agent and no robotic menu.
There is Tencent Hy Realtime, an AI that listens, understands, and answers in a voice that sounds almost human. If the customer asks it to, it might even sing.
Hy Realtime is the new real-time voice model from Tencent’s Hunyuan team, announced on November 13, 2026. It is already live inside the company’s Yuanbao assistant and its own customer service lines, and Tencent says a public API on Tencent Cloud will follow soon.
What Tencent Hy Realtime Can Do
Hy Realtime is built for conversation that happens in real time, not for typed prompts. It listens to spoken words, understands what the speaker means, follows instructions, and replies with natural-sounding speech, fast enough to feel like talking to a person.
According to Tech360TV, the model improves on the basics that matter most in a voice assistant: speech understanding, instruction following, and human-like expression. Those may sound like small steps, but they are the difference between an assistant you tolerate and one you actually enjoy talking to.
The bigger change is what Tencent calls agentic tool calling. Instead of only answering from what it already knows, Hy Realtime can take action during a conversation. It can decide for itself when a spoken question needs a web search, run that search, and bring the answer back without being told to.
That moves it from a simple responder to something closer to a worker. A voice model that only talks is a novelty. A voice model that goes and gets things done is a tool.
A Voice Model That Decides When to Search
Most voice assistants still wait for you to ask the right question in the right way. You have to know what they can do, and phrase your request to match. Hy Realtime flips that around: the model itself judges whether your question needs fresh information from the internet.
Ask it about something that happened this week, and it should recognise that its training data cannot answer you. It searches the web on its own, reads what it finds, and folds the result into its spoken reply. The user never sees a search box, and never has to ask twice.
This is the same direction the whole voice AI race is moving in. ElevenLabs’ Scribe v3 Medical, which we covered earlier this week, is building voice agents that take notes and trigger follow-up actions. OpenAI and Google are pushing their own assistants toward doing tasks, not just answering questions.
Tencent’s version stands out because it is already answering real customer calls. The model is live in Yuanbao voice calls and in Tencent’s own customer service, which means it is being tested on impatient, real-world callers, not just in a demo room.
It Can Sing Too
The most eye-catching feature is also the most unusual. Hy Realtime can sing. Ask it for a song, and it will perform one, with the kind of phrasing and emotion that makes a voice feel alive rather than assembled.
The singing was co-designed with an AI team at Tencent Music Entertainment, known as TME Lingyin. That partnership makes sense: Tencent Music runs some of China’s biggest music platforms, and its engineers know how a sung line should breathe, rise, and land.
Singing may sound like a party trick, but it is a serious test of a voice model. A spoken sentence forgives a lot. A song does not. Pitch, timing, and feeling all have to line up, or the result sounds wrong in a way anyone can hear.
If Hy Realtime can sing convincingly, it suggests the underlying voice engine is far more expressive than the flat, slightly robotic voices most of us associate with AI assistants.
Where You Can Try Hy Realtime Today
You do not need to wait for the Tencent Cloud API to hear the model. It is already available in the latest version of the Yuanbao app, Tencent’s AI assistant, where you can pick the voice called Chuyu in a voice call and talk to it directly.
There is also a web demo hosted on Tencent Cloud, where anyone can try a conversation in the browser. That is a useful way for developers and curious users to judge the model with their own ears before committing to anything.
Tencent has not given a date or a price for the coming Tencent Cloud API. For now, the company is describing it as planned for the near future, which means developers who want to build with Hy Realtime will have to wait a little longer.
That staged rollout is a familiar pattern in China. Alibaba’s Qwen4-Max, which we covered earlier this week, also arrived first as a limited preview before wider access. Tencent appears to be following the same playbook: prove the model in its own products, then open it to everyone else.
The Benchmark Claims and the Catch
Tencent says Hy Realtime scores well in its own testing. The company reports gains in internal blind tests and on several public voice benchmarks, describing its results on those public tests as leading the field.
That claim comes with an important caveat: the numbers are Tencent’s own, and the public reports do not include sample sizes or test settings. Until independent researchers run their own evaluations, the benchmark results should be treated as the company’s claim, not as a settled fact.
The announcement also leaves out the details technical users usually want first: no parameter count, no training data description, no standalone model card, and no formal API document yet. The public record is mainly media restatements of an official WeChat post from the Hunyuan team, plus the live demo.
That is worth remembering when the headlines sound definitive. Hy Realtime is a real product, already in real use, but the full technical picture has not been published. The singing and the searching are demonstrated. The rest is still on trust.
What Comes Next for Tencent Hy Realtime
Tencent has already sketched out where the model goes from here. The team says the next steps include duplex interaction, which would let the model listen and speak at the same time, more like a real conversation where both sides can interrupt.
Longer-term personalised memory is also on the roadmap. A voice assistant that remembers your preferences, your past conversations, and the way you like things done becomes far more useful than one that starts from zero every time you call.
The third item on the list is coordination between foreground and background agents: the voice you talk to could quietly hand tasks to other AI agents working behind the scenes, then report back when the work is done.
None of these features has a date yet. But the direction is clear: Tencent wants Hy Realtime to grow from a voice that talks into a voice that manages work on your behalf.
Why Tencent Hy Realtime Matters
Voice AI is becoming the next big battleground. Text chatbots are now commonplace, but voice is where the next billion interactions will happen: in cars, in customer service lines, in smart glasses, and in apps used by people who would rather talk than type.
China’s AI labs are moving fast in this space. Alibaba’s Qwen team, MiniMax, and now Tencent’s Hunyuan team are all shipping voice and multimodal models at a pace that would have been unthinkable two years ago.
Hy Realtime also shows how Tencent plans to use its own scale. It has the Yuanbao assistant, a giant customer service operation, and Tencent Music, each giving the model a real place to work and real users to learn from.
For now, the model is a Chinese-language story with a global echo. Whether it becomes a serious rival to the voice models from OpenAI, Google, and ElevenLabs will depend on what happens next: the API launch, the independent benchmarks, and the features Tencent has promised.
But one thing is already clear. The voice assistant that only answers questions is on its way out. The one that searches, sings, and gets things done has arrived, and it is already picking up the phone.
Source: Tech360TV



