What Is Cartesia?Cartesia 是什么？
Cartesia is a developer-first voice platform whose whole pitch is one number: latency. Its text-to-speech model family is Sonic — the current model is Sonic-3.6, a stable snapshot released 27 August 2026 and backwards compatible with Sonic-3.5, covering 44 languages after adding Odia and Urdu. It streams audio over REST or WebSocket as text is pushed, so a voice agent can start speaking before the full sentence is generated, and it supports voice cloning, voice localisation across accents, and controllable delivery (speed, emotion and pronunciation). Alongside Sonic sit Ink for speech to text (Ink-2 at 3 credits per second of audio) and Line for managed voice agents, with SDKs for Node, Python and Go and native LiveKit and Vapi integrations. One versioning detail worth knowing: the alias sonic-3.6 tracks the latest stable snapshot, while a dated model ID such as sonic-3.6-YYYY-MM-DD pins one that never changes — pin the date in production. Note the model-name divergence: some August 2026 documentation still lists Sonic-3.5 as the documented stable family, while October 2026 comparisons and Cartesia's own pricing page use Sonic-3.6; this page uses Sonic-3.6 as current and treats Sonic-3.5 references as the previous generation.Cartesia 是开发者优先的语音平台，整个卖点就一个数字：延迟。它的文本转语音模型家族是 Sonic——当前模型为 Sonic 3.6，2026-08-27 发布的稳定快照、向后兼容 Sonic 3.5，新增奥里亚语与乌尔都语后覆盖 44 种语言。它通过 REST 或 WebSocket 在文本推入时即流式返回音频，语音智能体不必等整句生成完就能开口，还支持嗓音克隆、跨口音本地化，以及语速、情绪与发音的可控表达。Sonic 之外还有转写的 Ink（Ink-2 每秒音频 3 积分）与托管语音智能体的 Line，提供 Node、Python、Go SDK 及 LiveKit、Vapi 原生集成。一个值得知道的版本细节：别名 sonic-3.6 跟随最新稳定快照，而 sonic-3.6-YYYY-MM-DD 这类带日期的模型 ID 则锁定永不变化的版本——生产环境请锁定日期。注意模型名分歧：部分 2026-08 文档仍把 Sonic 3.5 列为文档化稳定家族，而 2026-10 横评与 Cartesia 定价页均用 Sonic 3.6；本页以 Sonic 3.6 为当前代，把 Sonic 3.5 视为上一代。
The latency claim needs the same plain treatment. Cartesia markets sub-90 ms model latency for Sonic, and on its own path it is genuinely fast — but that is a model-side number. An independent benchmark measured the earlier Sonic-3 at about 188 ms median time to first audio end to end, and about 351 ms at P90 in Coval's August 2026 measurements. Both numbers are true; they measure different things. If the product is a voice agent, a tutor or a game character, budget against the benchmark and test on your own network. If the product is narration, dubbing or a studio voiceover, latency barely matters and the comparison changes entirely: on this site, ElevenLabs wins on voice catalogue size, language coverage and creator workflows, while Murf AI wins on the studio timeline for finished, aligned voiceovers — Cartesia wins only when the audio is generated live, inside a product, for each user.延迟宣称也需要同样直白的说法。Cartesia 为 Sonic 宣传的亚 90 毫秒是模型侧延迟，在它自己的链路上确实很快——但那不是端到端数字。独立实测中，上一代 Sonic-3 的端到端首音频中位数约 188 毫秒，Coval 2026-08 实测 P90 约 351 毫秒。两个数字都对，只是测的东西不同。如果产品是语音智能体、陪练或游戏角色，就按实测预算、在自己的网络上测一遍。如果产品是旁白、配音或录音室级画外音，延迟几乎不重要，比较就完全变了：本站 ElevenLabs 在嗓音库规模、语言覆盖与创作者工作流上占优，Murf AI 在成品对齐画外音的录音室时间轴上占优——Cartesia 只在音频为每个用户实时生成于产品内部时才胜出。
Verdict: buy Cartesia when the audio is live and the pause is the product — a voice agent, a tutor that must not be talked over, a game character that answers in character. Its $5 Pro plan is the cheapest credible way to test a real-time voice product with a commercial licence, and its credit math is honest once you convert to minutes: about 133 minutes on Pro, about 1,667 on Startup, about 10,667 on Scale. Do not buy it for narration or dubbing, where ElevenLabs' catalogue and Murf's studio beat it, and do not promise a client sub-90 ms end to end — promise what you measured on your own path. Pin the dated Sonic snapshot, watch the concurrency caps (2 / 3 / 5 / 15 concurrent TTS requests from Free to Scale), and treat the free plan as a prototype bench, not a launch plan.
实在的结论：只有在音频是实时的、停顿本身就是产品时才买 Cartesia——语音智能体、不能被打断的陪练、要即兴应答的游戏角色。$5 的 Pro 档是带商用许可测试实时语音产品最便宜的可信入口；换算成分钟后它的积分账是实在的：Pro 约 133 分钟、Startup 约 1667 分钟、Scale 约 10667 分钟。别为旁白或配音买它——那是 ElevenLabs 的嗓音库与 Murf 的录音室的领地；也别向客户承诺端到端亚 90 毫秒——承诺你在自己链路上测到的数字。锁定带日期的 Sonic 快照，盯住并发上限（免费到 Scale 的 TTS 并发请求为 2 / 3 / 5 / 15），把免费档当原型台、别当上线方案。
— 中国吧 AI Tools Hub Editorial Team
Who Should Use It?谁适合用？
✓ Good Fit For适合以下场景
- Voice agents, live tutors and game characters where latency is the UX语音智能体、实时陪练、游戏角色：延迟就是体验本身
- Developers shipping conversational products on WebSocket streaming用 WebSocket 流式交付对话产品的开发者
- Teams that need a cloned voice in production from 10 seconds of audio需要用 10 秒音频做出生产级克隆嗓音的团队
✕ Not Ideal For不适合以下场景
- Narration, audiobooks and dubbing where a studio workflow matters more旁白、有声书、配音：录音室工作流更重要的场景
- Anyone who needs a large stock voice catalogue out of the box开箱即要庞大现成嗓音库的人
- Mission-critical deployments without a fallback voice provider没有备用语音供应商的关键业务部署
Getting Started快速开始
1. Start free: 20,000 credits, instant cloning, playground voice test1. 先免费：2 万积分、即时克隆、演练场试音2. Measure TTFA on your own network; pin sonic-3.6-YYYY-MM-DD2. 在自己的网络上测首音频延迟；锁定带日期的 sonic-3.6 快照3. Convert credits to minutes (750-800/min), then pick the tier — checkout total wins3. 把积分换算成分钟（每分钟 750–800）再选档——以结算价为准Key Features核心功能
- Sonic-3.6: Real-Time TTS for Live ConversationSonic 3.6：为实时对话而生的 TTS — 44 languages, sub-90 ms advertised model latency, WebSocket streaming so speech starts before the sentence finishes.44 种语言、官方称亚 90 毫秒模型延迟、WebSocket 流式输出，句子没说完就能开口。
- Voice Cloning: Instant and Professional嗓音克隆：即时版与专业版 — Instant cloning from about 10 seconds of audio on every plan; professional voice cloning from Startup, with 2 slots on Startup and 4 on Scale.各档均可约 10 秒音频即时克隆；专业克隆从 Startup 起，Startup 2 个名额、Scale 4 个。
- Line: Managed Voice AgentsLine：托管语音智能体 — Calls at $0.06 per minute plus $0.014 per minute for a Cartesia-provided phone number, with prepaid agent dollars equal to the plan price.通话 $0.06/分钟、Cartesia 电话另加 $0.014/分钟，每档含等额于档位价的智能体预付金。
- Ink: Speech to TextInk：语音转文字 — Ink-2 at 3 credits per second of audio, about $0.39 per hour on the Scale plan per Cartesia's own pricing page.Ink-2 每秒音频 3 积分，按 Cartesia 定价页口径 Scale 档约 $0.39/小时。
- APIs, SDKs and Agent IntegrationsAPI、SDK 与智能体集成 — REST and WebSocket APIs, Node / Python / Go SDKs, native LiveKit and Vapi integrations, and a dated-snapshot pinning scheme for production.REST 与 WebSocket API、Node / Python / Go SDK、LiveKit 与 Vapi 原生集成，生产可用带日期快照锁定版本。
Pros & Cons优缺点
✓ Pros优点
- Fastest credible real-time voice path for agents, tutors and games语音智能体、陪练、游戏角色最可信的实时语音路径
- Honest credit math: 1 credit per character, minutes easy to compute积分账实在：1 积分/字符，分钟数好算
- Unused credits roll over up to twice the monthly amount未用积分最高结转至月额 2 倍
- Free plan prototypes a full voice agent without a card免费档不用绑卡就能做出完整语音智能体原型
✕ Cons缺点
- Advertised sub-90 ms is model-side; end-to-end benchmarks are slower宣传的亚 90 毫秒是模型侧；端到端实测更慢
- Old annual prices ($4 / $39 / $239) still circulate and mislead旧年付价（$4 / $39 / $239）仍在流传、会误导
- Free plan has no commercial licence; concurrency caps force upgrades免费档无商用许可；并发上限会逼着升级
- Younger vendor than the hyperscalers — pair with a fallback比超大厂年轻的供应商——关键业务请配备用厂商
Use Cases应用场景
Popular ways people use it.常见用法：
📞 Voice Agents That Do Not Get Talked Over不会被用户打断的语音智能体
Support, booking and triage agents that answer fast enough to hold a turn.客服、预约与分诊智能体：答得够快，才守得住话轮。
🎓 Live Tutors and Language Practice实时陪练与语言练习
Tutors that listen, respond and correct in the same breath, across 44 languages.边听边答边纠错的陪练，覆盖 44 种语言。
🎮 Game Characters That Improvise会即兴应答的游戏角色
NPCs that speak generated lines in character instead of canned recordings.NPC 不再放录音罐头，而是按角色即兴说出台词。
♿ Real-Time Accessibility Narration实时无障碍播报
Live captions read aloud and spoken navigation where waiting breaks the task.实时字幕朗读与语音导航：等不起的场景。
Similar AI Tools相似 AI 工具
Related pages on this site: ElevenLabs for the broadest voice catalogue and creator suite, and Murf AI for finished, timeline-aligned voiceovers. If the job is narration rather than live conversation, start there; if the job is a live agent, Cartesia is the benchmark to beat. This batch's 3D-generation sibling is Tripo.本站相关页：ElevenLabs（最广的嗓音库与创作者套件）与 Murf AI（成品对齐画外音的录音室时间轴）。如果工作是旁白而非实时对话，从那两家起手；如果工作是实时智能体，Cartesia 就是要打败的标杆。本批 3D 生成方向的同批页为 Tripo。