Tencent Hunyuan unveils Hy ASR 3.0 preview, bringing context-aware speech recognition to Yuanbao

On August 4, Tencent Hunyuan officially released Hy ASR 3.0 preview, a new-generation speech recognition model built on the language understanding of its latest large language model, Hy3. It fuses high-precision speech recognition with deep semantic understanding, delivering gains across general recognition, context awareness, multi-scenario robustness and dialect coverage, and producing accurate, coherent transcriptions that better match user intent in complex real-world inputs.

On multiple open benchmarks, Hy ASR 3.0 preview keeps word error rates (WER) around 3% across languages — 3.34% for Mandarin, 2.62% for English and 3.12% for Cantonese — leading its competitors. On self-built evaluation sets, it also posts low WERs for general recognition, dialect recognition, context understanding and hard acoustic conditions such as high noise and whispers.

Tencent highlights four capability upgrades: more accurate general recognition for dialects and mixed Chinese-English speech with fewer character errors; better understanding of user intent through contextual homophone correction; easier adaptation to professional scenarios via hotword injection for brand names, people's names and industry terms; and greater stability in complex environments, with targeted optimization for high-noise and whispered speech.

The gains come from architecture, data scaling and post-training working together. Hy ASR 3.0 preview uses an efficiency-focused MoE architecture with the Hy3 base model, plus a self-developed unsupervised speech encoder trained on tens of millions of hours of unlabeled speech to extract high-quality acoustic representations.

The team jointly trained the speech encoder with the LLM on tens of millions of hours of multi-source speech data covering dialects, accents and acoustic environments, refined through a high-quality data pipeline. Multi-stage capability injection, an SFT data system spanning 10 major dialect regions and over 20 second-level sub-regions, and multi-stage reinforcement learning for general transcription accuracy, any-context understanding and long-tail scenarios then sharpened each ability while reducing misrecognition.

Hy ASR 3.0 preview is already available as an API on the Tencent Cloud website, targeting use cases such as intelligent customer service, content understanding and voice search. Tencent's Yuanbao assistant, which co-developed the model, has launched it first — users can press and hold to talk and get dialect recognition, contextual error correction and stable transcription in complex environments, free of charge — with products like WorkBuddy rolling out next.