--- title: "Tencent's Hy ASR 3.0 Understands What You Meant, Not Just What You Said" subtitle: "Tencent's new ASR model plugs directly into Hy3's language brain, hitting ~3% WER across Mandarin, English, and Cantonese while understanding context, not just words." url: "https://alphasignal.ai/news/tencent-s-hy-asr-3-0-understands-what-you-meant-not-just-what-you-said" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Tencent Hy" company_url: "https://alphasignal.ai/company/tencenthunyuan" published: "2026-08-04T09:58:55.000Z" updated: "2026-08-04T12:08:11.771Z" type: "News" topics: - "Audio" - "Benchmarks" subtopics: - "Speech To Text" upvotes: 454 read_time: "5 min" source_url: "https://x.com/tencenthunyuan/status/2084579829303615497" comments: 0 --- # Tencent's Hy ASR 3.0 Understands What You Meant, Not Just What You Said > Tencent's new ASR model plugs directly into Hy3's language brain, hitting ~3% WER across Mandarin, English, and Cantonese while understanding context, not just words. Most speech recognition systems do one thing: turn audio into text, word by word, and stop there. [Tencent Hunyuan's Hy ASR 3.0 preview](https://x.com/tencenthunyuan/status/2084579829303615497) takes a different approach. By wiring its ASR pipeline directly into the language understanding capabilities of Hy3 -- Tencent's latest large language model -- it produces transcripts that reflect what you **meant**, not just what you said. Hy ASR 3.0 preview leverages the language comprehension capabilities of Hy3, combining high-precision recognition with deep semantic understanding. The practical result: it can provide more accurate and coherent transcriptions by combining context, achieving an evolution from word-by-word transcription to semantic understanding. ## The numbers On open benchmarks, word error rates reach 3.34% for Mandarin, 2.62% for English, and 3.12% for Cantonese, achieving around 3% overall with impressive multi-dimensional performance. That puts it in competitive territory with the current open-source ASR field, where NVIDIA's Canary Qwen 2.5B currently tops the Hugging Face Open ASR Leaderboard with 5.63% WER -- though that benchmark is English-focused and not directly comparable. Tencent also reports strong results on its internal evaluation sets, covering general recognition, dialect recognition, context understanding, domain-specific vocabulary, and complex acoustic scenarios like high noise and whispered speech. ## Four things it actually does better The team called out four specific capability upgrades over the previous generation: - **General accuracy:** It has been significantly improved in general recognition, context awareness, multi-scenario robustness, and dialect coverage. This includes Mandarin, dialects, and code-switched Mandarin-English speech. - **Context-aware correction:** The model uses surrounding context to disambiguate homophones -- words that sound identical but mean different things. This is a particularly hard problem in Mandarin, where tonal ambiguity is common in fast or informal speech. - **Hotword injection:** It supports injecting custom vocabulary -- brand names, product names, people's names, and industry jargon -- so businesses can get accurate transcription of domain-specific terms without retraining the model. - **Noisy environment robustness:** Tencent specifically optimized for high-noise environments, whispered speech, and low-volume audio, keeping WER stable where other models degrade. ## How it's built The architecture is a MoE (Mixture-of-Experts) design -- a setup where only a subset of the model's parameters activate for any given input, keeping inference costs manageable while maintaining a large total parameter count. On the speech side, the team built a proprietary unsupervised speech encoder trained on tens of millions of hours of unlabeled audio, enabling it to extract rich acoustic features from complex real-world conditions. The key design decision is the base model upgrade to Hy3. Hy3 is a fast-and-slow-thinking fused MoE language model with 295 billion total parameters, 21 billion activated parameters, and support for up to 256K context. Plugging ASR into that backbone means the transcription layer inherits Hy3's context modeling and semantic reasoning -- the same capabilities that make Hy3 competitive on coding and agent tasks now help the ASR system understand what a speaker is trying to say, not just decode phonemes. ## Where it fits right now Hy ASR 3.0 preview is already live on Tencent Cloud and available via API, targeting use cases like intelligent customer service, content understanding, and voice search. Tencent's Yuanbao assistant was the first product to integrate it -- users can long-press to speak and immediately experience dialect recognition, context-based error correction, and stable transcription in noisy environments, with the feature available for free. WorkBuddy and other Tencent products are rolling out support as well. The practical use cases where this matters most: - Call center quality inspection and agent assist, where noisy phone audio and industry jargon are the norm - Meeting transcription, where speakers switch languages mid-sentence or use company-specific terminology - Voice search and voice input in consumer apps, especially across Chinese dialect speakers - Content moderation and media subtitling pipelines that need semantic coherence, not just raw word output ## The real shift here The broader trend this release reflects is the convergence of ASR and LLMs into a single pipeline. Models like NVIDIA's Canary Qwen introduced a Speech-Augmented Language Model (SALM) architecture combining ASR with LLM capabilities, pairing a FastConformer encoder with an LLM decoder. Hy ASR 3.0 takes a similar approach but leans harder into the language model side -- using Hy3's full semantic reasoning rather than a lightweight decoder. The implication for anyone building voice products: a ~3% WER headline number matters less than whether the model understands context. Correcting a homophone or preserving a brand name correctly can be the difference between a usable transcript and one that needs manual review. That's the problem Hy ASR 3.0 is directly targeting -- and it's the right problem to be solving in 2026. The model is available now through [Tencent Cloud's ASR service](https://www.tencentcloud.com/products/asr). Pricing details for the 3.0 preview tier have not been separately published from the existing Tencent Cloud ASR pricing structure. The underlying [Hy3 language model](https://github.com/Tencent-Hunyuan/Hy3-preview) it builds on is open-source under Apache 2.0, though the ASR system itself is offered as a cloud API. ## Sources - [Tencent Hunyuan Hy ASR 3.0 preview announcement (Twitter/X)](https://x.com/tencenthunyuan/status/2084579829303615497) - [AIbase: Hy ASR 3.0 preview details and benchmark breakdown](https://www.aibase.com/news) - [163.com: Hy ASR 3.0 preview full technical coverage (Chinese)](https://www.163.com/dy/article/L3GKO8DU0511AQHO.html) - [Tencent Cloud ASR product page](https://www.tencentcloud.com/products/asr) - [Tencent Hunyuan Hy3 official release page](https://www.tencent.com/en-us/articles/2202386.html)