[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-voxtral-3-voice-agent-sdk":3,"topics-all":38,"news-related-bbedb317-c221-43e3-b9a4-61798ed17b55":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},"bbedb317-c221-43e3-b9a4-61798ed17b55","Voxtral 3:Mistral 语音智能体全栈,一轮对话不到一秒","Mistral 发布 Voxtral 3 语音家族:三款转写模型覆盖 50+ 语言,实时档延迟低于 500 毫秒;TTS 支持 35+ 语言 50+ 方言,13 美元\u002F百万字符。语音智能体 SDK 把听、想、说串成流式管线,单轮不到一秒,可经 SIP 接打电话,全栈可自部署到自有基础设施。","语音智能体至今没有像聊天机器人那样普及,卡点很直白:它要求模型同时做对四件事——跨语言跨口音听准、结合上下文想明白、带着情绪说得自然、还要在通话里连续调用工具把事办完。Mistral 把这四件事打包成 Voxtral 3:三款语音转写模型、一款 TTS 模型,加上一个语音智能体 SDK。\n\n## 转写:三档模型,最低 0.35 美分一分钟\n\n转写侧一次发了三款。Voxtral Small Transcribe 3 定位最高精度,官方称其多语言转写与说话人分离达到业界先进水平,在 FLEURS 45 语言宏平均测试中词错率 4% 出头,超过 Whisper、Scribe v2 和 Gemini 3.5 Transcribe。Voxtral Mini Transcribe 3 是 4B 参数小模型,batch 转写定价 0.0035 美元\u002F分钟,是全家族最低价。第三款 Voxtral Mini Transcribe Realtime 3 面向智能体场景,35+ 语言实时转写,延迟低于 500 毫秒,原生支持端点检测,让智能体知道用户说完了没。\n\n比较实用的是 context biasing:可以把领域术语、产品名、说话人名字作为关键词喂给模型。官方在 Earnings 22 上的测试里,Voxtral Small 的关键词召回率从 81.4% 拉到 94.6%,对术语密集的呼叫中心,这比微调专用模型便宜得多。\n\n## TTS:35+ 语言 50+ 方言,13 美元一百万字符\n\nVoxtral Mini TTS 3 支持 35+ 语言和 50+ 方言,普通话还区分北京、四川口音,定价 13 美元\u002F百万字符,官方称低于 OpenAI TTS-1、Deepgram、ElevenLabs Flash 和 Gemini Flash。它的卖点是\"隐式情感\":模型自己从文本上下文推断该用什么情绪,不需要额外指令。发音控制支持音素提示,官方对比里,Voxtral 1 会把航班时间「21:00pm」念成「21,000pm」,Voxtral 3 则能正确处理。第三方 Voice Code Bench(BeSimple AI)也把它放在质量\u002F价格曲线左上角。\n\n## SDK:一条不到一秒的流水线\n\n语音智能体 SDK 基于 LiveKit,把 Voxtral Mini Transcribe Realtime 3(听)、Mistral Small(推理与工具调用)、Voxtral TTS(说)串成一条流水线,官方称单轮响应不到一秒。智能体可以在通话中核身份、查账单、订预约,靠 function calling 和检索打通业务系统;浏览器和 App 走 WebRTC,也能通过 SIP 直接接打电话。按分钟计费:配 Mistral Small 推理 0.04 美元\u002F分钟,换更大模型 0.08 美元\u002F分钟。\n\n## 卖点其实是\"搬得回家\"\n\n模型能力之外,Voxtral 3 真正的差异化在部署:欧盟托管的 API,或者整套自部署到自有基础设施,音频处理与对话留存自己说了算。每一级都是独立模型,LLM、音色、语言都可以单独换。对受监管的金融、医疗呼叫中心,这张主权牌可能比跑分更有说服力——管线能整体搬回自己机房的选项,目前仍然少见。\n\n跑分是官方自评、评测 judge 也是模型,这些数字不妨打个折看。但 0.35 美分一分钟的转写加上 13 美元一百万字符的合成,把语音智能体的成本砸到了呼叫中心能认真算账的水平。当\"听-想-说-干活\"被打包成一条按分钟计费的流水线,剩的问题只有一个:你的业务系统准备好被打电话的 AI 调用了吗?\n\n参考:[Mistral 官方公告](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fvoxtral-3\u002F);The Decoder 同题报道。","https:\u002F\u002Fmistral.ai\u002Fnews\u002Fvoxtral-3\u002F","2436174c-644b-4a65-9a98-e7a3b705569a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"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},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"81cffbe8-06c6-4a83-bab6-686c6550ee2e","en","Voxtral 3: Mistral's Voice Agent Stack, Sub-Second Turns","Mistral's Voxtral 3: three ASR models (from $0.0035\u002Fmin, sub-500ms realtime), TTS in 35+ languages at $13\u002F1M chars, and a sub-second voice agent SDK.","Voice agents have yet to spread like chatbots, and the bottleneck is simple: a reliable one must do four things at once — listen accurately across languages and accents, reason about intent in context, speak naturally with emotion, and execute multi-step actions over connected systems. Mistral has packaged all four into a single family: Voxtral 3, comprising three speech-to-text models, one TTS model, and a voice agent SDK that chains listening, reasoning, tool calling, and speaking into one streaming pipeline.\n\n## Transcription: three tiers, from 0.35 cents per minute\n\nThe transcription side ships three models at once. Voxtral Small Transcribe 3 targets highest accuracy; Mistral claims state-of-the-art multilingual transcription and diarization, with a macro-average word error rate around 4% on FLEURS across 45 languages, outperforming Whisper, Scribe v2, and Gemini 3.5 Transcribe — note these rankings come from the company's own benchmarks. Voxtral Mini Transcribe 3 is a 4B-parameter model priced at $0.0035\u002Fmin for batch transcription, the cheapest in the family. The third, Voxtral Mini Transcribe Realtime 3, is built for agentic use: real-time transcription in over 35 languages at sub-500ms latency, with native end-of-turn detection so the agent knows when the caller has finished speaking.\n\nParticularly practical is context biasing: you can feed domain terms, product names, and speaker names into the model. In Mistral's tests on Earnings 22, Voxtral Small's keyphrase recall jumped from 81.4% to 94.6%. For terminology-heavy contact centers in healthcare or legal, that is far cheaper than fine-tuning a dedicated model.\n\n## TTS: 35+ languages, 50+ dialects, $13 per million characters\n\nVoxtral Mini TTS 3 covers 35+ languages and 50+ dialects — Mandarin alone splits into Beijing and Sichuan accents — priced at $13 per million characters, which Mistral says undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. Its selling point is implicit emotion: the model infers the right emotional register from text context without extra instructions. Pronunciation control accepts phoneme hints; in Mistral's own comparison, Voxtral 1 would read a flight time of \"21:00pm\" as \"21,000pm,\" while Voxtral 3 handles it correctly — exactly the kind of detail (numbers, emails, proper names) where enterprise support agents fail. The third-party Voice Code Bench by BeSimple AI also places it in the upper-left, quality-per-dollar corner.\n\n## SDK: a sub-second pipeline\n\nThe voice agent SDK is built on LiveKit and chains Voxtral Mini Transcribe Realtime 3 (listening), Mistral Small (reasoning and tool calls), and Voxtral TTS (speaking) into one stream; Mistral says each turn completes in under one second. During a call the agent can verify identity, check a bill, or book an appointment using function calling, web search, and retrieval over your own systems. It runs over WebRTC in browsers and apps, and can place and answer calls over SIP. Pricing is per minute: $0.04\u002Fmin with Mistral Small for reasoning, $0.08\u002Fmin with a larger model.\n\n## The real pitch: you can take it home\n\nBeyond raw model quality, Voxtral 3's differentiation is deployment: an EU-hosted API, or a full self-hosted stack on your own infrastructure, with control over where audio is processed and whether conversations are retained. Every stage is a separate model, so the LLM, the voice, and the language can each be swapped independently. For regulated finance and healthcare contact centers, that sovereignty argument may matter more than benchmarks — OpenAI's GPT-Live and Microsoft's streaming speech models are all sold as APIs, but a complete voice pipeline you can run in your own server room remains rare.\n\nTake the scores with a grain of salt: the benchmarks are vendor-run and the eval judge is itself a model. But transcription at 0.35 cents a minute plus synthesis at $13 per million characters pushes the cost floor of voice agents down to where contact centers can seriously run the numbers. Once \"listen, reason, speak, act\" becomes a per-minute pipeline, one question remains: is your business system ready to be called by an AI on the phone?\n\nReference: [Mistral announcement](https:\u002F\u002Fmistral.ai\u002Fnews\u002Fvoxtral-3\u002F); [The Decoder](https:\u002F\u002Fthe-decoder.com\u002Fmistral-ai-launches-voxtral-3-a-full-voice-agent-stack-that-listens-thinks-and-talks-in-real-time\u002F)","voxtral-3-voice-agent-sdk","2026-10-21T23:11:24Z","2026-10-21T23:11:26.981031Z","2026-10-21T23:11:26.981044Z",true,"agent",39,[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},"96a256d8-cdca-4730-a566-a27985c5517c","微软语音三连发:流式转写百毫秒出稿,TTS 延迟 150 毫秒","microsoft-mai-transcribe-2-streaming","2026-10-02T19:09:40+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"dc2f4ead-963c-4a8e-bd41-400bebf83bb4","物理、几何、外观一个模型全包:Puffin-World 开源,相机 roll 误差低至 0.26°","puffin-world-native-3d-world-states","2026-09-06T19:09:41+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"6062d551-9068-4a9f-ae8e-4e99269cd838","Muse Voice Transcribe 发布:流式转写、20+ 说话人分离、端点检测,Meta 全塞进一个模型","meta-muse-voice-transcribe-streaming-asr","2026-09-05T13:11:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"630ed9ae-699e-4115-8e75-33196ea6db28","MiniMax Music 3 开源:8B+0.6B 双 LLM 写五分钟完整歌,8GB 显存能跑","minimax-music3-open-weights-architecture","2026-08-29T13:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"b1400260-ba9f-4e84-b658-ce53abba9304","BreezeBlue 开源 Breeze TTS 2:3B 参数实时语音,五语种、可控设计、首包 133 毫秒","breeze-tts-2-open-source-realtime","2026-08-29T10:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"fdf05035-e02e-4954-8beb-c697ecc7975a","Cohere Parse 发布:$1.5 每千页的文档解析模型,ParseBench 79.2 超 Mistral OCR 4","cohere-parse-document-parsing-model","2026-08-28T13:30:00+00:00"]