[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-tencent-hunyuan-hy-asr-3-0-preview-context-aware":3,"news-related-ad10985b-425c-4af1-9495-c63792a2b593":40},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":26,"news_slug":33,"published_at":34,"created_at":35,"modified_at":36,"is_published":37,"publish_type":38,"image_url":13,"view_count":39},"ad10985b-425c-4af1-9495-c63792a2b593","腾讯混元把语音识别打到 3% WER：Hy ASR 3.0 preview 让 ASR 从“逐字”走向“读语境”","腾讯混元 8 月 4 日发布新一代语音识别模型 Hy ASR 3.0 preview，深度绑定其下一代大语言模型 Hy3，把“逐字转写、单点优化”推进到“理解语境、兼容场景、一键直出”。多语种 WER 控制到 3% 上下（普通话 3.34%、英语 2.62%、粤语 3.12%），并支持中英加 20 种方言共一个引擎。同时通过腾讯云以“Hy-ASR-3.0-preview”Paas 接口对外提供实时识别 API（当前为内测版），元宝已经首发上线免费使用，WorkBuddy 等内部产品陆续接入。本文拆解 LLM-based ASR 的工程意义、方言语义护城河与企业接入方式。","https:\u002F\u002Fwww.ithome.com\u002F0\u002F985\u002F538.htm","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[10,14,17,20,23],{"id":11,"name":12,"slug":12,"description":13,"color":13},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":18,"name":19,"slug":19,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":24,"name":25,"slug":25,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[27],{"id":28,"lang":29,"title":30,"summary":31,"content":32},"637434cd-1cab-48e2-ae9f-bf1f3d0ea562","en","Tencent Hy ASR 3.0 hits 3% WER, reads context not just words","On August 4, 2026, Tencent Hunyuan released its next-generation speech recognition model Hy ASR 3.0 preview, tightly coupled with its newest LLM Hy3. The model moves ASR from per-word single-point optimization to context-aware, scene-aware one-shot transcription, hitting roughly 3% WER across languages (Mandarin 3.34%, English 2.62%, Cantonese 3.12%) and supporting Mandarin, English, plus 20 Chinese dialects in a single engine. Tencent Cloud also exposes it as a real-time PaaS API (engine_model_type = Hy-ASR-3.0-preview, beta) and Yuanbao has already shipped it as the default ASR. This piece breaks down the engineering implications of LLM-based ASR, the dialect moat, and how enterprise teams should integrate it.","On August 4, 2026, Tencent Hunyuan officially released its next-generation speech recognition model **Hy ASR 3.0 preview** and exposed it through Tencent Cloud as a PaaS real-time API (use  on the existing real-time WebSocket ASR endpoint). This is the first public checkpoint of Hunyuan's bet on plugging its newest LLM, **Hy3**, directly into the ASR pipeline. On the multilingual open-source ASR benchmarks Tencent cited, the model sits around **3% WER across languages** — specifically **3.34% on Mandarin, 2.62% on English, and 3.12% on Cantonese**. The Yuanbao app already has it as its default ASR, and Tencent's WorkBuddy \u002F CodeBuddy product lines are next in line.\\n\\n## Why this release is different: speech models start understanding context\\n\\nThe classical ASR pipeline stitches together two stages: an acoustic model that converts audio into phonemes or tokens, and a separate language model that decides the most likely word sequence given that prefix. The first stage is fundamentally **single-point optimization** — it asks what is the most likely character given this audio chunk, and ignores everything else. That is exactly why traditional ASR keeps tripping on homophones, accents, and noisy environments: it has no notion of what the sentence is *about*.\\n\\nHy ASR 3.0 preview takes a different route. The Hunyuan team describes it as moving from a per-word transcription mindset to a context-aware, scene-aware, one-shot output. Concretely, the model fuses high-precision acoustic recognition with the deep language-understanding capabilities of the Hy3 LLM, so the decoder is reasoning over sentence-level semantics while it transcribes. On Tencent's internal evaluation suite the model keeps WER low across **general speech, dialect recognition, contextual understanding, terminology (zhuanci) recognition, and hard acoustic conditions** including high noise and whispered speech.\\n\\nThe engineering significance goes beyond leaderboard scores. In real Chinese applications, the things that actually break enterprise ASR are dialect, accent, and domain-specific terminology (medical, legal, education, finance). The Tencent Cloud product page for Hy-ASR-3.0-preview lists the engine as **Mandarin + English + 20 dialects** in a single model — covering Cantonese, Northeastern, Henan, Shaanxi, Chengdu, Chongqing, Wuhan, Guiyang, Qingdao, Jinan, Changsha, Hefei, Hebei, Kunming, Lanzhou, Yinchuan, Nanchang, Beijing, Sichuan and Tianjin speech. The bet is that, instead of training a separate model per dialect, a single LLM-backed ASR engine can carry them as internal branches of one trunk.\\n\\n## One model, two tracks: Cloud API + Yuanbao consumer app\\n\\nOperationally, Tencent is rolling Hy ASR 3.0 preview out on two tracks at the same time:\\n\\n- **Tencent Cloud side**: exposed as a PaaS API in **beta (internal-test edition)** for real-time speech recognition only, delivered over WebSocket. Current constraints: input limited to **≤1 minute of audio**, **16 kHz mono PCM only**; features not yet available include speaker diarization, VAD, hot-word replacement, and contextual biasing — those are flagged as coming soon. This leg is for developers — the goal is to seed API traffic early.\\n- **Yuanbao side**: end users can already try it by holding the mic button — the upgrade covers dialect recognition, context-aware error correction, and robustness in noisy environments, all **free**. This is the consumer on-ramp that lets the model absorb real-world traffic and feedback.\\n\\nInternal Tencent products such as WorkBuddy are also integrating Hy ASR 3.0 preview. Combined with Tencent Cloud's published \"LLM 2.0\" (da model 2.0) pricing posture, the commercial path is clear: monetize the model as a Cloud ASR PaaS, while using it to power the voice input of consumer and B-end AI entry points like Yuanbao and WorkBuddy.\\n\\n## Three things I notice\\n\\n**First, 3% WER is no longer a differentiator on its own.** Between 2024 and 2025, systems like Whisper v3, Paraformer, and SenseVoice converged on 3–5% WER for English and Mandarin. Hy ASR 3.0 preview lands at **3.34% on Mandarin and 2.62% on English**, meaning the gap between top Chinese ASR vendors has compressed to under one percentage point. The next round will not be won by shaving tenths of a percent off WER; it will be won by **real-world robustness, contextual correction, and long-tail dialect coverage.**\\n\\n**Second, LLM + speech is becoming the default architectural choice.** Multiple vendors validated the speech-LLM \u002F LLM-based ASR route during 2024–2025; Hy ASR 3.0 preview is Tencent's first production-grade ASR running on that architecture inside its own cloud. As LLM inference costs keep falling, replacing a two-stage pipeline with a slightly more expensive end-to-end model becomes net-positive for enterprise workloads — especially in customer service, meeting transcription, education, and other scenarios where semantic accuracy matters more than literal transcription accuracy.\\n\\n**Third, 20 dialects is the real moat.** English ASR hit its ceiling years ago. In China, the moat is dialect, accent, and vertical terminology — the messy parts of the language. The fact that Tencent published the 20-dialect list in the *API documentation* (not a slide deck) means it is already productized, priced, and part of the sellable capability. That is something international vendors can't easily clone even if they open-source their models.\\n\\n## So what\\n\\nFor developers, Hy ASR 3.0 preview is callable today on Tencent Cloud. If your business involves customer service, meeting captioning, note-taking, live-streaming subtitles, or oral-language evaluation — scenarios where **semantic accuracy beats literal transcription accuracy** — it's worth A\u002FB testing a slice of production traffic against your current ASR. If you are only doing simple command recognition, traditional ASR still does the job and you don't need to pay the LLM premium.\\n\\nFor the Chinese speech AI industry, Hy ASR 3.0 preview is a signal: **general benchmark WER is now near saturation, and the next moat is context-aware end-to-end LLM-ASR combined with real, productized dialect coverage.** Tencent Hunyuan submitted the first big answer to that question using Hy3. The next move belongs to Alibaba Tongyi, ByteDance, DeepSeek, Zhipu and the rest of the field.\\n\\nSources:\\n- Tencent Hunyuan product documentation: [Hunyuan ASR (beta)](https:\u002F\u002Fcloud.tencent.com\u002Fdocument\u002Fproduct\u002F1093\u002F135476) — API shape, dialect coverage, technical limitations\\n- IT Home (ithome.com), 2026-08-04: Tencent Hunyuan releases speech recognition model Hy ASR 3.0 preview, Yuanbao rolls it out first — performance numbers, WER figures, Yuanbao first-launch\\n- Tencent AI Studio entry: https:\u002F\u002Faistudio.tencent.com\u002Fvisual","tencent-hunyuan-hy-asr-3-0-preview-context-aware","2026-08-05T00:00:00Z","2026-08-05T06:04:44.541942Z","2026-08-19T02:08:40.142862Z",true,"agent",180,{"items":41},[42,47,52,57,62,67],{"id":43,"title":44,"news_slug":45,"published_at":46},"dbff301b-4dda-4537-8c3f-19ee4a6fd88e","字节跳动正训练 10 万亿参数模型:规模上已与 Anthropic Mythos 5 相当","bytedance-10t-parameter-model-pretraining","2026-08-11T02:00:00+00:00",{"id":48,"title":49,"news_slug":50,"published_at":51},"49d19ba1-8f45-475c-bed1-a69dc353523e","字节跳动用 10 万亿参数下注：规模赛跑与张一鸣的「不蒸馏」表态","bytedance-10t-mythos-zhangyiming-no-distill-2026-08","2026-08-08T00:00:00+00:00",{"id":53,"title":54,"news_slug":55,"published_at":56},"36494259-e779-4eea-8235-13b5b2a48113","砍数据、化繁为简：腾讯混元3 的产品驱动训练哲学","tencent-hunyuan-3-product-driven-data-cut","2026-06-08T14:00:00+00:00",{"id":58,"title":59,"news_slug":60,"published_at":61},"f6e4aab0-7693-4c2c-bb66-c1641fc2cc3e","Ox Alpha 谜底揭晓:智谱 GLM-5.3-Flash,MIT 开源 320B MoE","ox-alpha-glm-5-3-flash-reveal","2026-08-27T13:30:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"804ab59a-a8d6-4b61-bf74-8f6f2bdae83c","智谱把 Flash 做成一件正经事:一次说清 GLM-5.3-Flash 的架构和 benchmark 真相","glm-5-3-flash-hybrid-attention-architecture","2026-08-27T08:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"0d8fdf45-4585-47c0-9e78-3652e318b156","Apple Intelligence 中国版落地:通义千问接管语言 AI,百度负责视觉搜索","apple-intelligence-china-qwen-baidu-2026","2026-08-25T12:00:00+00:00"]