Tencent Hunyuan launches HyRealtime realtime voice model with tool calls and singing

腾讯混元 · wechat · 2026-11-13

Tencent Hunyuan released HyRealtime, a new realtime voice model already live in Yuanbao voice calls and Tencent Customer Service, with API access coming to Tencent Cloud soon.

Highlights: better intent recovery from incomplete/rambling speech, stronger multi-turn instruction following and context management; new agentic tool-calling (deciding when to search web and unifying decision/tool/answer flow); and singing on demand via co-design with Tencent Music's TME Lingyin AI team, supporting natural-language song requests across artist, era, genre and language.

Technically, the team cut cross-modal IQ loss by rebuilding SFT data pipelines—raising text-to-speech capability retention from 21% to 69%—built tens of millions of hours of pretraining data across CT/SFT/OPD/RL stages, used multi-teacher cross-modal OPD plus dual-track reward RL with safety gating, and adopted low-latency streaming speech synthesis. Roadmap: full-duplex interaction, long-term personalized memory, and background agent interactions.

Related event: Tencent Hunyuan Unveils Realtime Voice Model HyRealtime(2 posts)→

Original post →

More from Models

Models channel →