Tencent Hunyuan ships Hy ASR 3.0 preview via cloud API
Tencent's new speech model reports about 3% word error rates in Mandarin, English and Cantonese, and is live on Tencent Cloud and in Yuanbao.
Tencent's Hunyuan team released a preview version of its next-generation speech recognition model, Hy ASR 3.0, on Tuesday, positioning it as a transcription system built on top of a large language model rather than a standalone acoustic pipeline.
What was released
According to Hunyuan's announcement, Hy ASR 3.0 preview borrows the language understanding of the team's Hy3 foundation model and uses it to resolve ambiguity that pure acoustic decoding leaves behind — homophones, proper nouns, domain jargon and speech that only makes sense given what came several sentences earlier. Tencent frames the shift as moving from character-by-character transcription toward a model that interprets the utterance in context and returns a cleaned, punctuated result in one pass.
On public benchmark sets, the team reports word error rates of 3.34% for Mandarin, 2.62% for English and 3.12% for Cantonese. On Tencent's own internal evaluations, it claims lower error rates than comparison systems across four axes: general recognition, dialect identification, contextual understanding and robustness in noisy or overlapping-speech conditions. Cantonese and broader dialect coverage is emphasised, an area where Chinese-market ASR products have historically been weakest.
The model is a commercial service, not an open-weight release. It is available through Tencent Cloud's API for use cases such as customer service, voice search and content indexing, and Tencent's consumer assistant Yuanbao has already switched it on, giving users free access to dialect recognition and context-aware correction.
Why it matters
Speech recognition has quietly become the contested layer beneath voice agents: an agent's tool calls are only as reliable as the transcript it acts on, and a misheard order number or address propagates through the entire chain. By routing recognition through an LLM's language prior, Tencent is making the same architectural bet as recent releases from Alibaba's Qwen ASR line and xAI's voice stack — that transcription accuracy is now bounded by semantics, not acoustics.
The pricing and rollout choice is also notable. Tencent is putting the model straight into Yuanbao at no cost while metering it through the cloud API, the same dual-track pattern Chinese vendors have used to seed developer adoption. With Cantonese and dialect handling near Mandarin-level error rates, the release also strengthens Tencent's hand in Hong Kong and Southeast Asian markets, where multilingual and code-switched speech is the norm rather than an edge case.