正文 · AI 翻译
IT Home reported on November 13 that Tencent Hunyuan released a new real-time voice model, Hy Realtime, today. The model performs well in basic conversational capabilities such as speech understanding, instruction following and humanlike expression, and adds agentic tool calling and singing capabilities.
Hy Realtime is already available in products including Yuanbao voice calls and Tencent Customer Service. It will also become available on Tencent Cloud soon, with API services offered externally.
According to the company, Hy Realtime is still being refined. Its next priorities are improving full-duplex interaction, long-term personalized memory, and interaction between foreground and background agents to provide users with a better interactive experience.
The current version can be tried in the latest Yuanbao app by selecting the “Chuyu” voice in the voice call feature. Some capabilities can also be tried through the web demo.
https://rtcube.cloud.tencent.com/component/hunyuan/index.html
IT Home provides the official introduction below:
Hy Realtime performs well at understanding user intent and following instructions. In real-life conversations, users may speak unclearly, leave sentences unfinished, or change their minds and correct themselves mid-sentence. The model can use the conversational context to reconstruct the intent behind such incomplete input, significantly improving understanding accuracy. Its ability to follow instructions across multiple turns and manage context has also improved, allowing it to understand user intent consistently and keep conversations flowing even during longer exchanges.
In addition to basic conversation, Hy Realtime can call tools. Based on the conversational context and user intent, it can accurately decide whether to perform a web search, bringing search decisions, tool calls and answer generation together.
Alongside a substantial improvement in listening comprehension, Hy Realtime's spoken expression has also been upgraded. It surpasses the previous-generation model across sound quality, pronunciation, prosody and emotion. Changes and progression of emotion within conversations are more natural, its handling of prosody and intonation is more varied, and its ability to empathize with users has improved further.
Through close co-design with Tencent Music TME's Lingyin AI team, Hy Realtime has gained the ability to sing, integrating song requests in natural language and expressive singing into conversations, taking interaction from “able to speak” to “able to sing.”
Users can request songs using combinations of singer, album, genre, language, release year and popularity, and can continue selecting based on earlier conversation. For example, when a user says, “Play an English folk song from between 2010 and 2020,” the model can draw on the context across multiple turns to understand the request, find a matching song and sing it.
For singing performance, the Hunyuan team has continued refining expressiveness, naturalness and musicality, giving the singing more nuanced emotion, more natural articulation and transitions between breaths, and better coordination of melody and rhythm. In internal expert A/B tests, Hy Realtime showed significant advantages across core experience measures.
Advertising disclosure: External links in this article, including but not limited to hyperlinks, QR codes and codes, are provided to convey more information and save time in evaluating options. Results are for reference only. All IT Home articles containing external links include this disclosure.