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)→
More from Models
- Google AI Mode now enforces usage limits, users surprised by rate-limit popup — elvano124 · 2026-11-22
- Sol 6.2 now refuses reverse engineering that Sol 6.1 happily did for a month — cardak98 · 2026-11-22
- Anthropic's record ~$81B IPO, GPT-6.2 pricing, Qwen4-Max and Copilot Autopilot: week in review — ksraj1001 · 2026-11-22
- Reddit rant: Frontier AI labs are training models to imitate lazy engineers — RossPeili · 2026-11-22
- MBZUAI open-sources PAN 2, a 29B-parameter general-purpose world model — LinusEkenstam · 2026-11-22
- Inkling Mini turns out surprisingly fluent in Russian, sparking data recipe questions — teortaxesTex · 2026-11-22