[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-earlybird-29m-predictive-streaming-asr":3,"topics-all":38,"news-related-fbce7a2f-800a-4af2-9e30-932ab3ed2a53":67},{"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},"fbce7a2f-800a-4af2-9e30-932ab3ed2a53","语音识别学会抢跑:29M小模型提前90ms出词","流式语音识别历来要等一个词说完才输出,普遍晚90到400毫秒。Earlybird把29M参数小模型放在随机截断的音频上训练,提前写出正在说的词、顺带猜下两个词:六个测试集上中位词都在说完前正确输出,有声书里69%的词提前命中,置信度保持校准。代价是更频繁收回输出,背景人声下准确率明显落后大模型。","和语音助手对话,它总慢半拍——问题往往不在大模型,而在识别器:流式语音识别要等一个词说完才敢输出,通常晚 90 到 400 毫秒。对语音 Agent 来说,这段时间本可以用来发起搜索、预取数据、组织回复。arXiv 新论文 Earlybird 把这个默认设定掀了:一个 29M 参数、4-bit 权重只有 16.8 MB 的小模型,在说话人说完之前就把词打出来,有声书测试集中位词提前 90 毫秒正确输出,每个词还附带一个校准过的置信度。\n\n## 把「等」改成「猜」\n\nEarlybird 的做法很直接:把 Moonshine Tiny(34M,MIT 许可)的词表砍到英语语音实际用到的 19,007 个 token,模型缩到 29M 参数、精度几乎不丢(每个测试集 WER 差距不到 0.4),再做 4-bit 量化感知训练,权重文件 16.8 MB。\n\n训练的核心技巧是喂「随机截断的音频」:以 80 毫秒为步长随机掐断,要求模型写出此刻所有「到期」的词——只听了一半的词也要完整拼写——再附上下两个词作为猜测。由于相邻截断点的目标序列只会向前扩展,模型从训练里就没学过「收回已经说出口的话」。两个细节值得注意:蒸馏损失只在教师模型拿对的 token 位置施加,否则猜词会学得犹豫不果断;给编码器补三帧静音「垫帧」,把中位提前量从 50 毫秒推到 90 毫秒。\n\n## 六个测试集,全部提前\n\n论文在六个带词级时间戳的测试集上对比了四个发布模型(全部用各自官方代码跑流式):\n\n| 测试集 | Earlybird | Moonshine Tiny | Whistle | Whisper base |\n|---|---|---|---|---|\n| 有声书 clean | **-90 ms** \u002F 2.3 | +270 \u002F 3.9 | +400 \u002F 3.1 | +90 \u002F 2.3 |\n| TED 演讲 | **-68 ms** \u002F 5.4 | +298 \u002F 6.8 | +505 \u002F 7.3 | +125 \u002F 4.8 |\n| 财报电话 | **-46 ms** \u002F 18.4 | +309 \u002F 21.6 | +535 \u002F 17.8 | +142 \u002F 13.3 |\n\n六个集全部提前输出(另有声书 other -60、公共演讲 -66、合成语音 -73);有声书上 69% 的词在说完前就正确命中,对比模型只有 3%-25%。在 Open ASR Leaderboard 八个英文集上,4-bit 版平均 WER 12.6,八个集全部优于 Moonshine Tiny 官方流式运行时(14.4),与同规模的 Whistle(12.2)大致打平;全精度版 11.4,比大一号的 Whisper tiny.en(12.0)还略准。算力账同样能看:单 CPU 线程上每 80 毫秒音频平均耗时 39 毫秒,约两倍实时。最实用的是置信度:对正在说的词,期望校准误差 3.7-5.5,置信度 ≥0.9 的词 91%-96% 是对的——下游 Agent 拿到的是「词+可信度」,可以自己定阈值决定信不信。\n\n## 代价与边界\n\n抢跑不是白拿的。提前输出意味着更频繁收回:有声书上每个词平均收回 1.29 次,远低于 Moonshine 官方运行时(2.07),但远高于死等的 Whistle(0.16);自发语和嘈杂环境升到每词 2-6 次。背景人声是最大短板:10 dB 嘈杂下 WER 32.0,Whisper base 只有 8.9。模型仅支持英语,域外命令语音(SLURP)上校准明显退化(ECE 8.8)。而「堆大模型」并不能自然解决延迟:为流式而生的 Nemotron 0.6B(464 MB)中位词反而 +230 毫秒;70 倍大小的 Parakeet 0.6B 每 240 毫秒重解码一遍能做到 -10 毫秒,但那是以 Earlybird 一半的错误率换来的。\n\n## 所以呢\n\n这篇论文真正动的不是精度,而是「等词说完」这个很少被当成设计变量的架构默认。从 +270~+309 毫秒到 -90 毫秒,相当于凭空送给下游 Agent 近 400 毫秒的思考时间——在语音交互里,这未必比把 WER 再降一个点便宜。配合校准置信度,识别器第一次把「信不信」的决定权交给了调用方。模型以 MIT 许可开源在 [HuggingFace](https:\u002F\u002Fhuggingface.co\u002Fsahilmahendrakar\u002FEarlybird-29M),16.8 MB、单核 CPU 两倍实时,端侧语音 Agent 的拼图又齐了一块。当识别器开始抢跑,下一个该优化的延迟,是不是轮到「说话人开口之前」了?(论文:[arXiv 2611.14421](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14421))","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14421","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"c4455801-0273-477d-9e92-00b4addfa439","asr",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"26b850e6-e60d-463b-8c20-6f876e219458","en","Earlybird: 29M ASR Predicts Words Before You Finish Them","The 29M Earlybird, trained on randomly cut audio, prints words ~90 ms before they end; 69% of audiobook words land early with calibrated confidence.","Talk to a voice assistant and it always lags half a beat — and the problem is usually not the LLM but the recognizer: streaming speech recognition waits until a word is finished before emitting it, typically 90–400 ms later. For a voice agent, that is time it could spend starting a search, prefetching records, or planning a reply. A new arXiv paper, Earlybird, flips that default: a 29M-parameter model with 4-bit weights of just 16.8 MB prints words before the speaker finishes them — on the audiobook test set the median word is output correctly 90 ms early — and attaches a calibrated confidence to every word.\n\n## Turning \"wait\" into \"guess\"\n\nThe recipe is straightforward. Earlybird takes Moonshine Tiny (34M, MIT license), cuts its vocabulary down to the 19,007 tokens English speech actually uses, shrinking it to 29M parameters with almost no accuracy loss (within 0.4 WER points on every test set), then runs quantization-aware training to 4 bits for a 16.8 MB weight file.\n\nThe core training trick is feeding randomly truncated audio: cut at random 80 ms steps, the model must write every word that is \"due\" at that moment — spelling half-heard words in full — plus a guess of the next two words. Because targets at successive cut points only ever extend one another, the model is never trained to take back what it has said. Two details matter: the distillation loss is applied only at token positions where the teacher is right (otherwise guessed words learn to be indecisive), and padding the encoder with three silence frames pushes the median earliness from 50 ms to 90 ms.\n\n## Six test sets, all early\n\nThe paper compares four released models (each run with its own official code in streaming mode) across six test sets with per-word timestamps:\n\n| Test set | Earlybird | Moonshine Tiny | Whistle | Whisper base |\n|---|---|---|---|---|\n| Audiobooks clean | **-90 ms** \u002F 2.3 | +270 \u002F 3.9 | +400 \u002F 3.1 | +90 \u002F 2.3 |\n| TED talks | **-68 ms** \u002F 5.4 | +298 \u002F 6.8 | +505 \u002F 7.3 | +125 \u002F 4.8 |\n| Earnings calls | **-46 ms** \u002F 18.4 | +309 \u002F 21.6 | +535 \u002F 17.8 | +142 \u002F 13.3 |\n\nIt is the earliest on all six sets (the others: audiobooks other -60, public speech -66, synthetic -73). On audiobooks, 69% of words are predicted correctly before they end, against 3–25% for the comparison models. On the Open ASR Leaderboard's eight English sets, the 4-bit release averages 12.6 WER — beating the official Moonshine Tiny streaming runtime (14.4) on all eight, roughly tying same-size Whistle (12.2), while the full-precision model reaches 11.4, slightly better than the larger Whisper tiny.en (12.0). The compute story holds up too: 39 ms on average per 80 ms audio step on a single CPU thread, about twice real time. Most practical of all is confidence: for words still being spoken, expected calibration error is 3.7–5.5, and words output at confidence ≥0.9 are right 91–96% of the time — downstream agents get \"word + trust score\" and can set their own threshold.\n\n## The costs and the edges\n\nPrediction is not free. Early output means more revisions: 1.29 take-backs per spoken word on audiobooks — far below the Moonshine runtime (2.07) but far above wait-it-out Whistle (0.16) — rising to 2–6 on spontaneous or noisy speech. Background chatter is the biggest weakness: at 10 dB, WER is 32.0 against 8.9 for Whisper base. The model is English-only, and calibration degrades on out-of-domain command speech (SLURP, ECE 8.8). And scaling up does not solve latency by itself: the purpose-built streaming Nemotron 0.6B (464 MB) lands at +230 ms median, while Parakeet TDT 0.6B — 70× Earlybird's size, re-decoded every 240 ms — reaches -10 ms but at half of Earlybird's error rate.\n\n## So what\n\nWhat this paper really moves is not accuracy but the architectural default of waiting for words to end — a variable rarely treated as a design choice. Going from +270~+309 ms to -90 ms hands downstream agents nearly 400 ms of free thinking time, which in voice interaction may be worth more than another point of WER. With calibrated confidence, the recognizer hands the \"trust it or not\" decision to the caller for the first time. The weights are open-sourced under MIT on [HuggingFace](https:\u002F\u002Fhuggingface.co\u002Fsahilmahendrakar\u002FEarlybird-29M) — 16.8 MB, twice real time on one CPU core — another piece of the on-device voice agent puzzle. Once the recognizer starts running ahead, is the next latency to optimize \"before the speaker even opens their mouth\"? (Paper: [arXiv 2611.14421](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14421))","earlybird-29m-predictive-streaming-asr","2026-11-11T13:30:00Z","2026-11-11T13:08:41.051496Z","2026-11-11T13:08:41.051511Z",true,"agent",205,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"09df4ac8-eede-4a9c-b488-0a577b589417","0.6B语音模型不生成答案,只回答是非题","marimba-typed-speech-decision-frozen-asr","2026-11-17T01:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"f08a17d3-2856-46ca-a104-27174e709b8a","TTS模型自己当编码器:参数省48%,三基准双胜","rse-tts-shared-reference-encoding","2026-11-15T19:40:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"6a099c03-4df1-434f-acbf-4a6a675f9ce7","神经草稿变慢变耗电:0.2毫秒查表草稿换来2.43倍","olifant-memory-draft-speculative-decoding","2026-11-14T09:20:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"aa23f73d-0db7-4b6e-a70e-1d72bc4f0127","严格因果ASR:Gabor字典卷积把流式状态钉在19MB","e-gaborformer-strictly-causal-streaming-asr","2026-11-07T11:20:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"9cfac3fd-f359-467e-94ba-f8ed3172d713","生成器自己当重排器:RAG 少驻留一个模型","glance-generator-reranker-shared-prefill","2026-10-31T13:11:54+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"e880f146-5f06-4b07-a4da-e6a7391ddcff","DeepSeek 开源 DeepInfer:生产推理引擎放出来,服务那一半还留在内部","deepseek-deepinfer-open-source-inference-engine","2026-10-22T03:10:00+00:00"]