Tencent’s Yuanbao can now sing on request during voice calls

Tencent introduced its real-time voice model Hy Realtime on November 13 and put it into calls in its Yuanbao AI assistant. The app’s official account said callers can ask for a song and hear the assistant sing it back. The update also changes how the assistant follows a conversation when a caller pauses, corrects a phrase or refers back to something they said earlier.

Yuanbao identifies one way to reach the new model: the voice named Chuyu (初语) in free conversation mode on the latest app. Voice calls themselves were already part of Yuanbao. Tencent’s Chinese App Store listing describes those calls and its QQ Music integration as separate features. The November update changes what the assistant can do within the call using the Chuyu voice.

A song request can start with a genre

Tencent’s Hunyuan team developed the singing feature with Tencent Music’s Lingyin AI team, according to its release explanation published by IT Home. A caller can combine criteria such as artist, album, genre, language and release period. In the company’s example, a caller names a genre and an era and leaves the choice of track to the assistant. A follow-up request can refine the choice using what was said earlier in the call.

Yuanbao already had a connection to recorded music. In its March earnings release, Tencent Music described the QQ Music integration as a way to stream songs through Yuanbao. It also described a Yuanbao-powered assistant inside QQ Music that could take requests in ordinary language and get users to digital albums. This week’s release adds a sung reply inside a Yuanbao conversation. Tencent’s public launch materials give examples of how to ask for a song and leave the full list of available tracks and the duration of a performance unspecified.

Search stays within the conversation

Web search is the tool the Hunyuan team describes for the new voice model. The company says the model uses the request and the conversation around it to decide whether to search. When it does call the tool, it can incorporate the result into its spoken reply. The existing Yuanbao app listing already describes search and voice input. The update brings the search decision into the live exchange.

Yuanbao illustrates the change with a mock interview. When a caller gets stuck, the assistant can draw on their earlier answers to offer a cue or continue with a follow-up question, the company says. Hunyuan also describes handling incomplete speech and corrections mid-sentence. Those examples concern what the model can use from the current call. A separate example from Yuanbao shows how the assistant can recap a topic if the caller loses their place. The Hunyuan team listed longer-term personalized memory as work for a future update.

A web demo now and a cloud API to follow

Tencent Cloud has a separate browser demo labeled Hy Realtime. Its public interface has options for captions and microphone settings. A search switch lets the assistant pull in outside information. The demo tells users to end the call before changing that setting. The page also asks for microphone access and displays a 10-minute limit per call.

The demo distinguishes between listening and preparing an answer. While the assistant is speaking, the page says a caller can interrupt by speaking or pressing a button. It displays a separate status when the microphone is muted and provides a notice when a call ends after a period of silence.

Hunyuan says Tencent has also put the model into its own customer service. The release describes that deployment without listing the types of inquiries it handles or which of the app’s features carry over. Developers looking to integrate Hy Realtime into their own services will get a separate route through Tencent Cloud, according to Hunyuan’s launch announcement.

Tencent Cloud already documents a voice conversation service that connects to several models, including a Tencent-built option. The existing guide describes how to connect those models. It gives no identifier for Hy Realtime. The launch materials leave the model’s API identifier, access requirements, launch date and price open.