[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ibm-granite-speech-5-turboctc-470m":3,"topics-all":38,"news-related-d72b4ca8-008c-4bc7-b527-e9d131d6a379":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"d72b4ca8-008c-4bc7-b527-e9d131d6a379","IBM Granite Speech 5.0:把 ASR 里的语言模型砍掉,470M 跑出 3.5 小时\u002F秒","IBM 开源 Granite Speech 5.0 TurboCTC,纯 470M 参数 encoder 架构跑出 12,600+ RTFx,H200 上 1 秒转写 3.5 小时音频;非商用版 WER 4.85%,Apache 2.0 版 5.00%。","## 事件:IBM 把 ASR 推回到 encoder-only\n\n2026 年 8 月 25 日,IBM 在 Hugging Face 社区博客上同步放出 Granite Speech 5.0 TurboCTC 的两块模型:granite-speech-5.0-470m-turboctc(Apache 2.0)和 granite-speech-5.0-470m-turboctc-nc(CC-BY-NC-SA-4.0)。两块都是 470M 参数,只在训练数据和许可证上有差异,核心架构完全一致([来源:IBM Granite HF 博客](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fibm-granite\u002Fgranite-speech-5-0-470m-turboctc))。\n\n这件事最反直觉的点在于:IBM 这次主动把 Granite Speech 家族里的语言模型和 LoRA 适配器砍掉了。前几代 Granite Speech 模型采用\"声学 encoder + projector + Granite LM + LoRA 适配器\"的三段式,新模型直接退回到纯 encoder + CTC。IBM 在博客里把这块拆得很清楚:\"The encoder-only design provides strong transcription performance, a small memory footprint of only 470M parameters, and over 20x faster throughput than previous Granite Speech models.\"\n\n砍掉 LM 的代价是显式的:模型放弃了语音翻译、keyword biasing 这些需要 LM 参与的能力。但 IBM 把收益的另一面摆在 OpenASR Leaderboard 上:在公开 short-form 测试集上,非商业版 WER 4.85%、Apache 2.0 版 WER 5.00%,聚合 RTFx 超过 12,600。这两个数字不是孤立的——前者保证了\"够准\",后者保证了\"够快\"。\n\n## 速度是怎么做到的:从 50 字符\u002F秒降到 12.5 token\u002F秒\n\nGranite 5.0 跟前代共享四个底层组件:16 个 Conformer block、第 8 层后接 self-conditioning、chunkwise attention 避免注意力随序列长度二次增长、CTC 损失函数。变化集中在输出端的 token rate。\n\n前代 encoder 每秒吐 50 个字符;新模型压到了 12.5 token\u002F秒,tokenizer 也换成了基于语音转录文本训练的 SentencePiece(NC 版)或 BPE(Apache 版)。从 100 帧\u002F秒的 log Mel spectrogram 压到 12.5 token\u002F秒,靠三段 2× 降采样:第一段是相邻 log Mel 特征向量的简单 reshape(这一段前代就在用),第二、第三段被嵌进堆栈最前面的两个 Conformer block,用 stride=2 的卷积做时间维下采样,同时把残差连接里的相邻位置做平均以匹配降采样后的时间步长。\n\n换句话说,不是更聪明的算法,是更稀疏的输出。token rate 砍到 1\u002F4,叠加纯 encoder 不再被 LM decode 拖慢,最终在 H200 上拿到 12,600+ RTFx——按 IBM 的说法,这意味着批推理 1 秒能跑完 3.5 小时以上的语音。同等准确度下,FFASR 远场榜单上 NC 版排名第五、Apache 版第九,两块同时还是榜单上速度最快的两个。\n\n## 为什么这件事对边缘设备更值得看\n\nGranite Speech 5.0 的副标题不是\"击败 Whisper\",而是\"ideal for speech-to-text tasks on edge devices\"。把 LM 砍掉之后,模型只剩 encoder,显存\u002F内存占用直接落到 470M 这一档;CTC 解码是贪心的,不需要 beam search 和大语言模型侧的 KV 缓存;Hugging Face 还配套上线了一个 Chrome\u002FEdge 才能跑的 WebGPU 流式 demo,意味着在浏览器里也能跑这套 470M 参数的 STT。\n\nIBM 在博客里还做了一件不太张扬但很实用的事:把训练数据配方完整列出来。两块模型共用 7 个公开数据集(MLS 44,600 小时、YODAS 8,900、CommonVoice-17 2,500、Librispeech 960、VoxPopuli 500、AMI 150、Earnings-22 100),NC 版额外接入 GigaSpeech 10,000 小时和 SPGI Speech 4,900 小时。除此之外还合成了 3 套:MLS\u002FYODAS\u002FCommonVoice\u002FVoxPopuli\u002FAMI 的多说话人拼接 2,000 小时、Earnings-22 拼接 500 小时,以及用 gpt-oss-120b \u002F gpt-oss-20b 生成文本、StyleTTS2 合成的 240 小时数字\u002F货币\u002F网址\u002F电话\u002F小数点专项语料——这一段直接展示了 IBM 怎么用大模型去造小模型的训练数据。\n\n## 给读者的判断:为什么这不只是一个 ASR 发布\n\n把这件事放到 2026 年下半年的趋势里看,信号其实很清楚:**通用大模型不适合所有任务,小而专的 encoder 在窄任务上正在反超**。Whisper、Canary、SenseVoice 这类端到端 STT 一直在堆参数、加 LM 加能力;IBM 反着走,把 LM 拆掉,只留 encoder,代价是丢掉翻译等\"看起来很酷\"的能力,换来的是 1\u002F20 的延迟和 1\u002F数十分之一的内存。\n\n对想在终端、浏览器、嵌入式设备上做实时转写的人来说,Apache 2.0 那块模型几乎是一个现成的生产级底盘:470M 跑得动 CPU,CTC decode 没有外部依赖,token rate 砍到 12.5 让流式输出更跟得上说话节奏。这也是为什么 IBM 把 WebGPU 流式 demo 放在 Hugging Face Spaces 上而不是论文图里——他们要让你打开浏览器就能体验\"砍掉 LM 之后到底有多快\"。\n\n所以如果你正在做边缘语音、车载语音、会议纪要、电话客服这类对延迟敏感、对翻译能力没要求的场景,Granite Speech 5.0 TurboCTC 是一个可以立刻拿去试的、许可证干净的 470M 选择;反过来,如果你需要语音翻译或 keyword biasing,前代 Granite Speech 仍然是 IBM 官方推荐的路线——这两件事在 IBM 的产品矩阵里是并行的,不是替代。\n\n参考链接:[Granite Speech 5.0 TurboCTC 模型卡(Apache 2.0)](https:\u002F\u002Fhuggingface.co\u002Fibm-granite\u002Fgranite-speech-5.0-470m-turboctc)、[WebGPU 流式 demo](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fibm-granite\u002Fgranite-speech-streaming-webgpu)。","https:\u002F\u002Fhuggingface.co\u002Fibm-granite\u002Fgranite-speech-5.0-470m-turboctc","653dda08-2edc-4d17-aeb2-56b0c88dd918",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"27fe0fb3-2b58-40d4-9152-fed94b935845","en","IBM Granite Speech 5.0: Drop the language model, 470M parameters transcribe 3.5 hours per second","IBM open-sources Granite Speech 5.0 TurboCTC: a 470M-parameter encoder-only ASR that hits 12,600+ RTFx on H200 (3.5 hours of audio per second under batched inference). On the OpenASR public leaderboard, the CC-BY-NC-SA-4.0 model scores 4.85% WER and the Apache 2.0 model scores 5.00% WER.","## Event: IBM Pushes ASR Back to Encoder-Only\n\nOn August 25, 2026, IBM posted Granite Speech 5.0 TurboCTC on the Hugging Face community blog and released two checkpoints side by side: `granite-speech-5.0-470m-turboctc` (Apache 2.0) and `granite-speech-5.0-470m-turboctc-nc` (CC-BY-NC-SA-4.0). Both are 470M parameters and differ only in training data and license; the core architecture is identical ([source: IBM Granite HF blog](https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fibm-granite\u002Fgranite-speech-5-0-470m-turboctc)).\n\nThe counterintuitive part is that IBM explicitly dropped the language model and LoRA adapters that previous Granite Speech models carried. Earlier generations used a three-stage stack — acoustic encoder + projector + Granite LM + LoRA adapters. The new models revert to a pure encoder + CTC. IBM states this directly: \"The encoder-only design provides strong transcription performance, a small memory footprint of only 470M parameters, and over 20× faster throughput than previous Granite Speech models.\"\n\nThe cost of dropping the LM is explicit: the new models give up speech translation and keyword biasing, the two capabilities that benefited most from having an LM in the loop. The benefit lands on the OpenASR Leaderboard: on public English short-form test sets, the non-commercial model scores an aggregate 4.85% WER and the Apache 2.0 model scores 5.00%, with aggregate throughput in excess of 12,600 RTFx. Accuracy and speed arrive in the same package rather than trading off against each other.\n\n## Where the Speed Comes From: 50 chars\u002Fsec to 12.5 tokens\u002Fsec\n\nGranite 5.0 shares four components with its predecessors: a stack of 16 Conformer blocks, self-conditioning after the 8th block, chunkwise attention to avoid quadratic scaling in sequence length, and CTC loss for training. The change concentrates at the output token rate.\n\nPrevious Granite Speech encoders emitted 50 characters per second. The new models drop to 12.5 tokens per second, with tokenizers trained on speech transcripts — SentencePiece for the NC model, BPE for the Apache model. Going from the 100-frames-per-second log Mel spectrogram down to 12.5 tokens per second relies on three stages of 2× subsampling. The first stage is a `reshape()` operation that stacks successive log Mel feature vectors (used by previous encoders too); the second and third stages are embedded in the first two Conformer blocks of the stack and use strided convolutions with `stride=2` for temporal downsampling. The convolutional residual path is temporally subsampled by taking the mean of successive pairs of positions so the residual still aligns with the downsampled time steps.\n\nIn other words, the speedup is not a clever new algorithm; it is a sparser output. Cutting the token rate to a quarter, combined with dropping the LM that bottlenecked decode, is what lands 12,600+ RTFx on an H200. By IBM's own framing, that is enough to transcribe 3.5 hours of speech in one second under batched inference. On the FFASR far-field leaderboard, the NC model ranks fifth in accuracy and the Apache model ranks ninth — and both are the two fastest models on that board at the same time.\n\n## Why This Matters More for Edge Devices\n\nGranite Speech 5.0's positioning is not \"beats Whisper\" but \"ideal for speech-to-text tasks on edge devices.\" Drop the LM and the model is just an encoder; memory footprint lands at 470M parameters; CTC decoding is greedy with no beam search and no language-model-side KV cache to manage. Hugging Face also hosts a WebGPU streaming demo that runs in Chrome or Edge, which means this 470M-parameter STT can run inside the browser.\n\nIBM also publishes the full training-data recipe. Both models train on seven public datasets (MLS 44,600 hours, YODAS 8,900, CommonVoice-17 2,500, Librispeech 960, VoxPopuli 500, AMI 150, Earnings-22 100). The NC version additionally pulls in GigaSpeech (10,000 hours) and SPGI Speech (4,900 hours). Three synthetic datasets go on top: 2,000 hours of multi-speaker concatenation from MLS\u002FYODAS\u002FCommonVoice\u002FVoxPopuli\u002FAMI, 500 hours of Earnings-22 concatenation, and 240 hours of utterances containing numbers, currencies, website names, phone numbers, addresses, and decimal points — generated by `gpt-oss-120b` or `gpt-oss-20b` and synthesized with StyleTTS2. That last block shows, in concrete terms, how a large general model was used to manufacture training data for a much smaller specialist model.\n\n## Reader Takeaway: Not Just Another ASR Drop\n\nRead against the second half of 2026, the signal is clear: general large models do not fit every task, and small specialist encoders are pulling ahead on narrow problems. Whisper, Canary, SenseVoice and similar end-to-end STT systems have kept scaling parameters and adding LM-side capabilities; IBM goes the other way and removes the LM. The price is losing translation and keyword biasing. The upside is roughly 1\u002F20 the latency and a fraction of the memory.\n\nFor anyone building real-time transcription on terminals, browsers, embedded devices, in-car stacks, meeting minutes, or phone-call analytics where latency matters and translation does not, the Apache 2.0 checkpoint is a near-drop-in production backbone: 470M fits on CPU, CTC decode has no external dependencies, and the 12.5 token\u002Fs rate keeps streaming output from falling behind the speaker. The WebGPU demo is the receipt — IBM is showing, not telling, what \"drop the LM\" buys you in the browser.\n\nIf you need speech translation or keyword biasing, the previous Granite Speech models remain IBM's recommended path. The two coexist in IBM's product matrix; they are not substitutes. For everyone else, Granite Speech 5.0 TurboCTC is the one to benchmark first.\n\nReferences: [Granite Speech 5.0 TurboCTC model card (Apache 2.0)](https:\u002F\u002Fhuggingface.co\u002Fibm-granite\u002Fgranite-speech-5.0-470m-turboctc), [WebGPU streaming demo](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fibm-granite\u002Fgranite-speech-streaming-webgpu).","ibm-granite-speech-5-turboctc-470m","2026-09-02T05:06:25Z","2026-09-02T05:06:37.056201Z","2026-09-02T05:06:37.056210Z",true,"agent",132,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"30fca629-bace-4832-9789-b44aa8c8989d","学生团队从零训出开源 7B 模型 ZGCM-1:数学推理硬刚 235B 前沿","zgcm-1-open-7b-foundation-model","2026-09-15T19:10:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"108af093-b226-4372-9cf0-77323ffc5456","小鹏 X-AuT 给语音大模型剪枝:音频塔砍 4 层,车载推理提速 21.4%","xpeng-x-aut-audio-encoder-pruning","2026-09-12T19:06:47+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"2b37a19b-1dde-4238-bef5-39b1d19157f1","OpenBMB 开源 MiniCPM5-2B:2B 端侧模型平均分超对比集 4B 级","openbmb-minicpm5-2b-on-device","2026-09-07T17:02:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"d941056b-c2e7-42e5-965a-a982c20b1169","Qwen3.8-Flash-Next 架构细节:Gated Residual 多分支残差 + QSA micro-block 稀疏注意力","qwen3-8-flash-next-cost-efficiency-architecture","2026-09-02T02:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"454286bd-cb8e-462e-8b33-1b4c77b27262","470M 语音模型 1 秒转写 3.5 小时:IBM 把 ASR 里的语言模型砍掉了","granite-speech-5-turbo-ctc-470m","2026-08-31T15:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"fb97a60d-69a1-4988-8de6-d1540ba63359","2.4B 参数读懂整页 A4:Cohere Labs 把最小的多模态模型挂上了 Apache 2.0","cohere-north-micro-vision-open-vlm","2026-08-18T13:30:00+00:00"]