Overview
End-to-end speech models share an old weakness that teams internally call getting dumber once they speak. The knowledge and reasoning a text model has thins out when it moves to voice: the sentences stay fluent, the thinking gets dull. Hunyuan aimed Hy Realtime squarely at that problem, released Nov 13, with a number attached: transfer retention up from 21% to 69%.
Read the number carefully. It measures how much capability survives when high-quality text data is transferred to speech tasks during post-training. It is not a benchmark score and not an accuracy figure. In plain terms, putting a reasoning model into a speech pipeline used to leave roughly a fifth of the intelligence usable; now close to seven tenths survives. The method was rebuilding the data pipeline that turns text into spoken phrasing, so the delivery sounds human without cutting the reasoning chain.
Of the two new capabilities, singing gets the attention and tool use matters more. Tool use lets the model decide for itself whether to search the web, folding the search decision, the tool call and the answer into a single pass rather than three separate stages. The cost of staging is most obvious in voice, where the user hears it as a few seconds of silence.
Singing came out of joint design with Tencent Music's Lingyin AI team. A request is not one sentence matched to one song: six dimensions can be combined, covering singer, album, genre, language, release year and popularity. The official example is an English folk song released between 2010 and 2020, which requires resolving the period, the language and the genre together. More useful is continuity, where a user narrows a previous result by switching language or shrinking the year range instead of describing everything again.
Speech understanding addresses the incomplete input of real conversation. Half sentences, mumbling, and self-corrections get reconstructed from context. That is hard to score but dominates how the product feels, because people rarely speak in complete utterances.
Availability needs stating plainly. It is live in the Yuanbao voice call feature and in Tencent's customer service, while the Tencent Cloud API is only a near-term plan, so developers cannot call it yet. To try it, pick the Chuyu voice inside the voice call feature of the latest Yuanbao app. The team says duplex interaction and long-term personalised memory come next, so this is not full duplex and handling interruptions remains limited.
The leaderboard section deserves caution. Internal expert blind reviews, self-built intelligence sets and several public benchmark scores all appear in the company's own materials, without third-party verification or reproducible configurations. They indicate direction, and should not be used to rank against other vendors.
Key Features
- Agentic tool use: The model judges from context whether to search, folding the search decision, the tool call and the answer into one pass, which removes the obvious pause in voice conversations.
- Singing with six-dimension requests: Co-designed with Tencent Music's Lingyin AI team, requests combine singer, album, genre, language, release year and popularity, and follow-up turns refine the previous result.
- Reconstructing incomplete input: Half sentences, mumbling and mid-sentence corrections are resolved against conversation context, and instruction following holds up across long exchanges.
- Text and speech generated together: Joint sequence modelling across text and audio means the model speaks while it writes rather than waiting for a full paragraph, cutting first-response latency for streaming interaction.
- More expressive prosody: The team reports gains over the previous generation in audio quality, pronunciation, prosody and emotion, with more natural emotional progression inside a conversation.
- Already in two product lines: Live in the Yuanbao voice call feature and Tencent customer service rather than a lab demo, reachable through the Chuyu voice in the Yuanbao app.
Use Cases
- Asking for current information in Yuanbao voice calls and letting the model decide whether to search
- Requesting songs by era, language and genre while driving or otherwise occupied
- Powering a Chinese-language customer service voice line that handles half sentences and corrections
- Building voice assistants for hardware and cars that need fast first response and tool access
Pros
- Transfer retention from 21% to 69% makes reasoning in voice conversations noticeably more usable
- Folding search decisions into the answer cuts the waiting that makes voice agents feel slow
- Six-dimension song requests with context carry-over match how people narrow choices when speaking
- Already serving traffic in Yuanbao and Tencent customer service rather than existing only as a paper
Pricing
Free for end users through the Chuyu voice in the Yuanbao app's voice call feature. The Tencent Cloud API is not open yet; the company says it will launch soon with API access, and no pricing has been published.
Summary
It addresses the old speech-model problem rather than adding a novelty. Retention rising from 21% to 69% means reasoning finally survives the move to voice, and tool use cuts the waiting. Singing will travel furthest as a demo and matters least among the three. Today there are only two entry points, so developers wait on the Tencent Cloud API.
Version History
- Release (2026-11-13): Tencent Hunyuan released Hy Realtime, a realtime speech model adding agentic tool use and singing. The model decides whether to search on its own, folding the search decision, the tool call and the answer into one pass; singing was co-designed with Tencent Music's Lingyin AI team and supports requests combining singer, album, genre, language, release year and popularity, with follow-up turns refining the previous choice. Post-training rebuilt the pipeline that turns text into spoken phrasing, lifting transfer retention from 21% to 69%. It is live in the Yuanbao voice call feature and Tencent customer service, with the Tencent Cloud API listed as a near-term plan and duplex interaction plus long-term personalised memory left for later.