[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-eleven-v4-tts-speech-arena":3,"topics-all":38,"news-related-02d83474-ba8d-402e-b4e3-be117be6a91d":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},"02d83474-ba8d-402e-b4e3-be117be6a91d","Eleven v4 发布:新架构登顶 TTS 盲测榜,推理延迟 100ms","ElevenLabs 发布新一代 TTS 模型 Eleven v4 与实时版 v4 Turbo:全新架构,支持 90 多种语言与高保真克隆,在 Artificial Analysis 盲测榜以 1319 Elo 登顶;Turbo 推理延迟中位数约 100 毫秒,瞄准实时语音 Agent。","9 月 28 日,ElevenLabs 同时发布两个语音模型:主打最高表现力的 Eleven v4,以及面向实时场景的 Eleven v4 Turbo。这代模型的目标不再是「把字念得像人」,而是「把文本演出来」——理解语气、情绪、节奏与上下文,同时保住说话人的音色。TTS 赛道正从合成质量竞赛,转向表现力与延迟的双轴竞争。\n\n## 新架构带来的三个直接升级\n\n[官方博客](https:\u002F\u002Felevenlabs.io\u002Fblog\u002Feleven-v4)明确说 Eleven v4 建立在全新架构之上。落到可验证的规格上,变化有三点:\n\n- **语言覆盖从 70+ 扩到 90+**:上一代 Eleven v3 支持 70 多种语言,v4 家族扩展到 90 余种,普通话与粤语都在支持列表里。\n- **单次生成上限翻倍**:Eleven v4 单次请求上限 1 万字符、约合 10 分钟音频,v3 是 5 千字符,长篇有声书受益最直接。\n- **声音克隆升级**:官方称 10 秒音频即可完成高保真即时克隆,效果超过 Multilingual v2 时代的专业克隆;v3 缺席的 Professional Voice Clones 也在 v4 回归。所有克隆需声音所有者验证同意,生成音频可用 AI Speech Classifier 检出。\n\n接口上,v4 提供 Text to Dialogue API 支持多人对话;Turbo 走 WebSocket 实时流。\n\n## 100 毫秒是语音 Agent 的胜负手\n\nv4 Turbo 的关键数字是推理延迟中位数约 100 毫秒,官方的原话是「比两个人对话之间的平均停顿更快」。在官方 9 月的 WebSocket 对比测试中(相同脚本、默认设置、剔除网络延迟),Turbo 首句语音中位耗时约 150 毫秒,Cartesia Sonic 3.6 为 262 毫秒,OpenAI GPT-4o mini TTS 为 814 毫秒。注意这是厂商自测口径,以第三方复测为准。\n\n更值得注意的是工程取向:官方明确表示 v4 Turbo 与自家 Agent 平台 ElevenAgents 是作为一个系统联合优化的。翻译过来就是模型与运行时绑定交付——语音 Agent 的延迟从来不只是模型问题,这正是 ElevenLabs 卡位的地方。\n\n## 盲测第一,定价试探\n\n在第三方盲测榜 Artificial Analysis 的 Speech Arena 上,Eleven v4 以 1319 的 Elo 位居第一,领先 Cartesia Sonic 3.6(1276)与 Google Gemini 3.8 Flash TTS(1267)。定价方面,据第三方汇总,Eleven v4 介绍期每千字符 2.2 美分、Turbo 为 1.1 美分,10 月 12 日之后恢复到 8 美分与 4 美分的原价。\n\n## 所以呢\n\n这轮发布给出的信号:质量差距已经被盲测榜量化,ElevenLabs 靠架构换代守住第一,但 Cartesia 与 Google 都贴在身后,没有代差。值得盯的是「模型 + Agent 运行时」的打包模式——当语音延迟进入 100 毫秒档,胜负手从模型参数转向系统级优化。留给竞品的问题是:你的 Agent 栈在哪里?\n\n来源:[ElevenLabs 官方博客](https:\u002F\u002Felevenlabs.io\u002Fblog\u002Feleven-v4)、[官方模型文档](https:\u002F\u002Felevenlabs.io\u002Fdocs\u002Foverview\u002Fmodels)、[Unite.AI](https:\u002F\u002Fwww.unite.ai\u002Felevenlabs-launches-eleven-v4-with-low-latency-turbo-variant)。","https:\u002F\u002Felevenlabs.io\u002Fblog\u002Feleven-v4","b359c914-654d-4630-bc3f-fba5eb3e8567",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"e5139045-7d07-4c36-b46f-5a85d6bb2d99","en","Eleven v4 tops TTS speech arena; Turbo at 100ms latency","ElevenLabs shipped Eleven v4 and v4 Turbo: new architecture, 90+ languages, Elo 1319 atop Artificial Analysis speech arena, ~100ms median latency.","On September 28, ElevenLabs released two voice models at once: Eleven v4, its most expressive text-to-speech model, and Eleven v4 Turbo, the realtime variant aimed at agents. The goal of this generation is no longer \"reading text like a human\" but \"performing it\" — interpreting tone, emotion, pacing and context while preserving the speaker's identity. The TTS race is shifting from pure synthesis quality to a two-axis contest of expressiveness and latency.\n\n## Three concrete upgrades from the new architecture\n\nThe [official announcement](https:\u002F\u002Felevenlabs.io\u002Fblog\u002Feleven-v4) states that Eleven v4 is built on an entirely new architecture. In verifiable specs:\n\n- **Languages grow from 70+ to 90+**: Eleven v3 supported 70+ languages; the v4 family covers 90+, including Mandarin and Cantonese.\n- **Double the per-request limit**: Eleven v4 handles 10,000 characters per request (about 10 minutes of audio) versus 5,000 for v3 — a direct win for long-form audiobooks.\n- **Better voice cloning**: instant clones now need just 10 seconds of audio and, per the company, outperform the Professional Voice Clones of the Multilingual v2 era; Professional Voice Clones, absent in v3, return with the model's full emotional range. Every clone requires verified consent from the voice's owner, and generated audio can be flagged by ElevenLabs' AI Speech Classifier.\n\nOn the API side, v4 ships via the Text to Dialogue API with natural multi-speaker dialogue, while Turbo streams over a WebSocket channel.\n\n## 100 milliseconds is the battleground for voice agents\n\nThe headline number for v4 Turbo is a median inference latency of about 100ms — in the company's words, faster than the average pause between two people talking. In ElevenLabs' September WebSocket benchmark (identical scripts, default settings, network latency removed), Turbo reached a median time to first speech of roughly 150ms, against 262ms for Cartesia Sonic 3.6 and 814ms for OpenAI GPT-4o mini TTS. Treat these as vendor-run numbers: the ordering is informative, and the exact figures need third-party replication.\n\nThe engineering direction matters more: ElevenLabs says it optimized v4 Turbo and its ElevenAgents platform together as one system. That means the model ships bound to the runtime — agent latency was never just a model problem, and architectures that stitch third-party TTS together are structurally disadvantaged. That is exactly where ElevenLabs is positioning.\n\n## First on the blind arena, and a pricing probe\n\nOn Artificial Analysis's blind-vote Speech Arena, Eleven v4 ranks first at Elo 1319, ahead of Cartesia Sonic 3.6 (1276) and Google Gemini 3.8 Flash TTS (1267). On pricing, per a third-party roundup, Eleven v4 costs 2.2 US cents per 1,000 characters during the intro period ($0.011 for Turbo), reverting to $0.08 and $0.04 after October 12.\n\n## So what\n\nThe signal from this launch: quality gaps are now quantified by blind arenas, and ElevenLabs held the top spot with an architecture change — but Cartesia and Google are right behind, with no generational gap. The thing to watch is the \"model + agent runtime\" bundle: once voice latency enters the 100ms class, the decisive factor moves from model parameters to system-level optimization. The question for competitors: where is your agent stack?\n\nSources: [ElevenLabs blog](https:\u002F\u002Felevenlabs.io\u002Fblog\u002Feleven-v4), [official model docs](https:\u002F\u002Felevenlabs.io\u002Fdocs\u002Foverview\u002Fmodels), [Unite.AI](https:\u002F\u002Fwww.unite.ai\u002Felevenlabs-launches-eleven-v4-with-low-latency-turbo-variant).","eleven-v4-tts-speech-arena","2026-09-29T13:07:54Z","2026-09-29T13:08:17.415987Z","2026-09-29T13:08:17.415996Z",true,"agent",667,[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},"d056f67b-7e0d-4e44-8d39-e31ea50deeae","Bonsai 2 27B 三元压缩:Qwen3.8 压到 5.9 GB,benchmark 留存 98.2%","bonsai-2-27b-ternary-qwen3-8-compression","2026-09-17T15:47:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"64b79fb5-b416-4b58-8402-0bc9ad22a27d","腾讯混元 Hy ASR 3.0 preview：MoE + 语音 Encoder 与 LLM 联合训练，把 ASR 推进到'理解语境'","tencent-hunyuan-hy-asr-3-0-preview-moe-speech-lm","2026-08-04T12:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"70b5b0d6-ce28-48e8-abe4-6a667a723c4e","xAI 把 Colossus 推到 2 GW:555,000 颗 GPU 撑起 Grok 4.6\u002F4.7 的万亿参数竞速","xai-colossus-2gw-grok-4-6-7-compute","2026-07-31T04:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"f4af1e4e-98a3-4810-831c-699ffe31ae73","马斯克公布 Grok 4.6\u002F4.7 路线图：1.5T\u002F2.1T 参数，SFT+RL 升级，8 月 7 日发行","grok-4-6-4-7-roadmap-1-5t-2-1t","2026-07-30T08:45:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"dea38861-2618-4468-9bad-a18eea96a818","Base44 Base 1：年入 1.5 亿美元的 vibe-coding 平台，终于把自己的 LLM 训出来","base44-base-1-vibe-coding-llm-launch","2026-07-29T06:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"b0183d10-bcfd-44ed-a178-a2c813f10b69","国家超算互联网AI社区上线Kimi K3:2.8万亿参数MoE一键调用,开源大模型有了国产算力底座","kimi-k3-scnet-platform-launch","2026-07-28T09:30:00+00:00"]