Tencent Hunyuan released Hy ASR 3.0 preview on August 4. QbitAI's full report says the model is based on Hy3 and combines speech recognition with semantic understanding. Tencent Cloud's documentation confirms that developers can call it through the real-time speech recognition WebSocket API by setting engine_model_type to Hy-ASR-3.0-preview.
The first numbers are WER figures
QbitAI and ITHome both cite the same open-benchmark figures: Mandarin WER of 3.34%, English WER of 2.62% and Cantonese WER of 3.12%. Tencent Cloud also lists 20 free concurrent channels and support for Mandarin, English and 20 Chinese dialects.
Those numbers describe different things. WER measures model output on benchmarks. Free concurrency and dialect coverage describe the current cloud product boundary. For developers, all three matter because API access, quota and city-level language coverage affect deployment cost.
The hook is context
Tencent's own wording says Hy ASR should move from "word-by-word transcription and point optimization" to "understanding context, adapting to scenarios and one-step output." In practical terms, the model must decide whether a homophone is a product name, a medical term or a project code by reading the surrounding speech.
QbitAI reports that the training recipe covers context, professional terms, acoustic environments and diverse speakers. Dialect data covers 10 major dialect regions and more than 20 sub-regions. Tencent also points to multi-stage reinforcement learning for general transcription, Any-context capability and long-tail acoustic scenarios.
The preview still has hard limits
Tencent Cloud says the current preview only supports real-time ASR, audio within one minute and 16k mono PCM input. Speaker separation, VAD, vocabulary replacement and noise-threshold settings are not yet supported. Context input and hot words are still marked as coming soon.
That makes the first fit clearer: short voice input, real-time captions, customer-service snippets, voice search and assistant commands. Long meetings, podcasts and multi-speaker interviews still need more product maturity.
Why it matters in China
The useful signal is not just another model launch. Voice is becoming an execution interface for agents, office assistants, customer-service bots and car systems. In China, Mandarin is only the baseline; dialects, accents, industry terms and noisy environments decide whether voice AI feels usable.
The next proof points are concrete: when context input and hot words open, whether the one-minute cap changes, whether more audio formats arrive, and whether Yuanbao and WorkBuddy show fewer real-world transcription mistakes.
Sources: QbitAI, ITHome, Tencent Cloud documentation, CocoLoop, CLS; sources verify Hy ASR 3.0 preview release timing, open-benchmark WER figures, 20 free concurrent channels, 20 dialects, preview input limits, API parameter and Yuanbao rollout.