Hy ASR 3.0 preview – Tencent Hunyuan's Next-Generation Speech Recognition Model Operational

Hy ASR 3.0 preview is Tencent Hunyuan's next-generation speech recognition model. Based on the Hy3 large language model, it combines high-precision speech recognition with deep semantic understanding. The model supports Chinese, English, Cantonese, and 10 major dialect regions, and provides contextual intelligent correction, hotword injection, and robustness in complex scenarios such as high noise and whispering. WER is kept at around 3% on open-source evaluation sets, while it comprehensively outperforms competitors on in-house evaluation sets.

Accurate general-purpose recognition: Supports Mandarin Chinese, English, Cantonese, and mixed Chinese-English speech, reducing misrecognized or omitted characters and cumulative errors in long audio.
Contextual semantic understanding: Uses Context to accurately capture the user's context, intelligently correct homophones, and eliminate semantic ambiguity.
Hotword injection support: Supports enhanced recognition of hotwords such as brand names, people's names, and industry terms, reducing the cost of business integration and maintenance.
Robustness in complex environments: Specifically optimized for acoustic scenarios such as high noise, whispering, and hushed speech to maintain stable recognition performance.
Broad dialect coverage: Supports 10 major dialect regions and more than 20 secondary subregions, including Northeastern Mandarin, Cantonese, Wu, and Southern Min.

Architecture upgrade: Adopts a MoE architecture and upgrades the base model to Hy3. Its self-developed unsupervised speech Encoder is trained on tens of millions of hours of speech data to extract high-quality acoustic representations.
Pre-training scaling: The Encoder and LLM are trained jointly with tens of millions of hours of multi-source speech data, including audiobooks, conversations, videos, and broadcasts. Multi-stage capability injection provides dialect understanding and contextual modeling capabilities.
Post-training enhancement: A high-quality SFT Recipe covering complex audio, Any Context, and general capabilities is developed. Multi-stage reinforcement learning is used to separately optimize general transcription accuracy, long-range contextual dependencies, and performance in complex long-tail scenarios.

Integrate it through the Tencent Cloud API, or open Tencent Yuanbao and press and hold the talk button to try it for free.

Upgraded semantic understanding: Based on the Hy3 large language model, it combines high-precision speech recognition with deep semantic understanding, evolving from word-for-word transcription to contextual understanding and one-click generation of polished output.
Comprehensive evaluation leadership: WER is kept at around 3% on open-source evaluation sets, while in-house evaluation sets show better performance than competitors in general, dialect, contextual, and complex acoustic scenarios.
Contextual intelligent correction: Uses Context to accurately capture the context, intelligently correct homophones, eliminate semantic ambiguity, and support long-range dependency modeling.
Rapid hotword adaptation: Supports hotword injection for brand names, people's names, industry terms, and other terms, reducing the cost of business integration and maintenance in specialized scenarios.
Dialect and complex-environment coverage: Supports 10 major dialect regions and more than 20 secondary subregions, while maintaining stable performance in complex acoustic environments such as high noise and whispering.

Intelligent customer service: In telephone and online customer service scenarios, Hy ASR 3.0 preview accurately recognizes brand names and business terms through hotword injection and enables real-time speech transcription, significantly improving service response efficiency.
Content creation: In the production of long-form audio content such as podcasts, meetings, and interviews, the model uses contextual semantic understanding to intelligently eliminate ambiguity and generate coherent, readable, high-quality text with one click, reducing post-editing costs.
Voice search: Supports queries spoken in dialects and accurately recognizes voice commands in noisy environments, effectively lowering the input barrier for users with different languages and accents.
Office collaboration: In real-time meeting transcription, the model maintains stable recognition even when handling low-volume inputs such as whispers and hushed speech, helping teams communicate efficiently and preserve information.
Smart devices: In interactions with IoT devices such as in-vehicle navigation systems and smart home appliances, the model is specifically optimized for high-noise acoustic environments to ensure robust speech recognition and a good user experience in complex scenarios.