腾讯混元Hy ASR 3.0 preview：让语音识别理解上下文Tencent Hunyuan Hy ASR 3.0 Preview: Speech Recognition That Understands Context

On August 4, Tencent Hunyuan officially released its new-generation speech recognition model, Hy ASR 3.0 preview. Built on the language-understanding capabilities of the latest large model, Hy3, Hy ASR 3.0 preview combines high-precision speech recognition with deep semantic understanding, delivering across-the-board improvements in general recognition, context awareness, multi-scenario robustness, and dialect coverage—producing transcripts that are more accurate, coherent, and closer to user intent in complex real-world inputs. It evolves from 'verbatim transcription with isolated optimizations' to 'understanding context, adapting to scenarios, and one-click output.'

On open benchmark sets, Hy ASR 3.0 preview leads overall. Its multilingual word error rate (WER) is around 3%: Mandarin Chinese 3.34%, English 2.62%, and Cantonese 3.12%—ahead of competitors. On Tencent's internal benchmarks covering general recognition, dialect recognition, context understanding, and complex acoustic scenes (high noise, whispering), WER also stays low and leads rivals.

From a user perspective, Hy ASR 3.0 preview focuses on four capability upgrades: (1) more accurate general recognition, reducing misrecognized and dropped characters across general speech, dialects, and Chinese-English mixing; (2) better intent understanding, using context to capture user meaning, intelligently correct homophones, and resolve ambiguity; (3) easier adaptation to professional scenarios via hot-word injection for brand names, people, and industry terms; and (4) greater stability in complex environments such as high noise, whispering, and soft speech.

The improvements come not from a single module but from joint gains in model architecture, data scaling, and post-training. Architecturally, it adopts an efficient MoE design and upgrades the base model to Hy3; on the audio side, a self-developed unsupervised speech encoder trained on tens of millions of hours of data extracts high-quality acoustic representations. Joint training of the encoder and LLM over multi-source speech data, plus multi-stage SFT covering 10 major dialect regions and 20+ sub-regions, and multi-stage RL, further cut errors in hard long-tail scenarios.

Hy ASR 3.0 preview is now available as an API on the Tencent Cloud site, usable in smart customer service, content understanding, voice search, and more. Yuanbao (Tencent's assistant) co-developed and launched it first—users can experience dialect recognition, context-aware correction, and stable transcription in noisy environments for free by long-pressing to talk; WorkBuddy and other products are being integrated.