On November 13, Tencent Hunyuan released Hy Realtime, a new real-time voice model pitched as able to talk, sing, and call tools. It is already live in Yuanbao voice calls and Tencent customer service, and it will soon arrive on Tencent Cloud as an API service. Just nine days earlier, on November 4, Hunyuan released Tencent Hunyuan Hy4, its first model with vision and computer operation; the real-time voice model follows close behind.
It catches half-finished sentences — and sings back
Hy Realtime handles messy input: mumbled words, sentences cut off halfway, mid-sentence corrections. It reconstructs what the user means from the conversation context, and its perception goes beyond speech-to-text into paralinguistic cues such as emotion, intonation, and background sounds. On expression, sound quality, pronunciation, rhythm, and emotion are all improved over the previous generation, Hy2-O, with more natural emotional shifts inside a conversation.
Two new capabilities stand out. The first is agentic tool use: the model decides from context and user intent whether a web search is needed, unifying search decisions, tool calls, and answer generation. The second is singing, built with Tencent Music's TME Lingyin AI team. Users can request songs by singer, album, genre, language, release year, and popularity, or continue choosing from earlier in the conversation — say "play an English folk song from 2010 to 2020" and the model understands the request, matches a song, and sings it.
The benchmark results need a closer look
In Hunyuan's internal expert blind review, Hy Realtime scored 2.52 overall, slightly below Doubao App's 2.56; its availability rate of 40.63 tied Doubao App for the top spot. Across four public benchmarks, it ranked first only on Wild Speech Bench, at 75.35, with a rival slightly ahead on the other three; on the in-house Hy-Eval-Audio3.0 it scored 39.33, below Gemini-3.8-Flash's 61.01. The clearest progress is against its predecessor: the overall score rose from 2.01 to 2.52, and availability from 25.78 to 40.63.
Anyone can try it now: open voice calls in the latest Yuanbao app, pick the "Chuyu" voice, and try it for free — including asking it to sing. Technically, Hunyuan raised the retention rate when transferring high-quality text data to voice tasks from 21% to 69%, the key to speaking like a person without losing intelligence. Next, the team will focus on full-duplex interaction, long-term personalized memory, and foreground-background agent interaction. Parameter count, pricing, and the API launch date have not been announced.
