Cartesia on X

A

arushh@Kcodess

built a voice gd room on seekho: four ai engineers argue, cross-question you and sit on one key fact. one llm call per line picks the speaker, code enforces the turns. cartesia streams voices in ~300ms, sarvam handles hinglish push-to-talk. ends with a scorecard of urs

J

Jimmy Mass@jimymass

Real app, fake backend. That's the trick behind the tutorial reels SpacePlanner makes every night, like this one on planning an electrical layout. Point a recorder at production and every take changes: new data, slow calls, random states. Mock the backend and the real UI does the same thing every time. The nightly setup: → cron at 22:17 UTC → Claude Code headless, with a backlog of topics → headless browser drives the live editor, backend mocked → screen recording, Cartesia voice-over, captions → frame check, then a summary by morning About 3,500 voice characters a night, across 6 reels. Some nights a topic gets skipped. The summary says so, and that note matters more than the reels. Repeatable input makes repeatable video.

H

Himesh Patel@HimeshPatel2121

added tools for live cricket scores, weather, and availability checks, and integrated cartesia TTS the full pipeline is now working end to end, and it actually speaks back clearly feels so good to see everything come together after building it from scratch can lookout on demo

M

Marc@MarcMil8

He creado LAURA 🎙️ Marcas una zona de la pantalla y, cuando aparece una frase nueva, o un texto, la lee en voz alta con tu propia voz clonada en Cartesia. Es una página web: no hay que instalar nada. Gratis y de código abierto. 👉 link

A

Akshay Ramaswamy@TheRealAk914

Such a fun video to film!! @cartesia is truly best in class in what they do, excited to keep pushing this industry forward together 🤝

M

Maxim Makatchev@maxipesfix

Driving parameterized models with voice has a lot of potential, for avatars and beyond. @JonPTaylor's pipecat-visemes does just that: it maps TTS audio onto mouth shapes in real time and streams them via @pipecat_ai's RTVI. The mapping can be sensitive to the TTS voice though. Some @cartesia voices say "who" with frequencies that suggest "ee", so the avatar smiles instead of rounding its lips. So I've added a calibration step: the voice reads 38 short words ("Heed. Tea. Bead. Knee. …", about 30 s of audio), and the analyzer learns where that voice puts each vowel. The demo below shows some cases where the differences are strong. gh repo: OpenHRIai/vrm-lipsync

负

负宇RL@RL_landing

Fish Audio 开源 350 亿参数语音模型 Drama 3，支持多角色同台与多轮讲戏重录 Fish Audio 开源了语音生成模型 Drama 3。它能在一个剧本里给最多 16 个角色配音，还能听你的修改意见重录上一句台词。模型权重已在 Hugging Face 公开，官网和 API 也能直接用。 现在的语音模型把字念准、把声音模仿像，已经不难了。但拿来做有声剧、游戏配音或语音 Agent，还会卡在三件事上。第一，指令是一次性的。语气录歪了，你没法对着刚才那条录音说“再温柔一点”，只能改提示词重抽。第二，一句话里只读错一个词，也得整句重录，前后原本好的语气全会变掉。第三，你让它笑一声或压低声音耳语，它常常演不出来，听着还是像在念稿。 模型演不出戏，根源在训练后期的打分规则上。过去的语音模型训练时，通常靠两个自动裁判打分。一个是语音转文字软件，检查有没有读错字。另一个是音色比对软件，检查声音像不像参考录音。这套规则天生排斥表演。演员一旦笑出声或叹气，转文字软件就会判它多读或漏读。演员一旦改成耳语或哭腔，嗓音变了，音色比对软件就会判它不像本人。Fish Audio 在论文里拿真人录音做过测试。真人用耳语念台词时，99.7% 的样本都被传统音色比对判为不如平淡朗读。带笑声念台词时，也有 93.9% 被判输。 Drama 3 干脆扔掉了所有现成的规则裁判。团队自己训练了一个专门听戏的打分模型。它直接看你的剧本要求和生成的音频编码，按真人评委的偏好打分。只要剧本里写了耳语或笑声，演出来的版本就能拿高分。官方测试显示，去掉转文字裁判后，模型并没有乱吞字。在中英文测试里，它第一次生成就出废片的比例反而降了一半左右。不过这样做也有代价。因为不再死磕刻板的音色分数，在只给一段参考录音的公开测试里，它的客观音色相似度只有 0.605。这个分数明显低于 OmniVoice 的 0.729 和 Qwen3-TTS 的 0.704。 为了听懂修改意见，Drama 3 把配音、改戏和局部补录，全装进了同一种多轮对话格式里。你可以直接用大白话在台词里加指令。比如写 [sighs] 让它叹气，或者用 <whisper> 包住半句话让它耳语。第一遍不满意，你把刚才的音频留在对话历史里，下一轮写“比上一条更轻柔温暖”，它就会照着重录。要是长句子里只改一两个词，它会把改动处前后的原录音夹在提示里，只重录中间那几秒。官方测试里，这种局部修补的中文读错率是 2.35% 到 3.69%，比同类接口 Cartesia Sonic-3 低了一半以上。在 24 种语言测试中，它的中、英、德等五种语言错误率也比 ElevenLabs v4 更低。但它的粤语错误率高达 22.70%，比竞品差不少。多轮改情绪的准确度，也还不如专门做语音编辑的 Step-Audio-EditX。 把多轮历史录音和整场戏塞进上下文，模型按理说会变慢。Drama 3 的总参数量涨到了 350 亿，是上一代 S2 的七倍多。但在单张 H200 显卡上，官方测出它的单流生成速度反而从每秒 177 帧提到了 271 帧。变快主要靠两处改动。第一，它用了混合专家结构，把模型拆成很多小模块，每次只唤醒 30 亿参数干活。专家权重还压缩成了占空间减半的 8 位格式。第二，它换掉了生成声音细节的方式。语音模型每生成一帧声音，都要算出 10 个代表音色细节的编码。显卡跑这种生成时，最卡脖子的不是计算，而是从显卡内存搬运权重的带宽。上一代 S2 用一个 4 层的小模块，在同一帧里循环跑 10 遍。每遍都要重读一次权重，一帧下来要跑 40 层、搬运 15.6 GB 数据。Drama 3 改用 13 层解码器，每一层专门负责一个编码，每帧顺着走一遍就行。这样一帧只需跑 13 层，搬运的数据量降到了 6.0 GB。单卡 H200 发出第一包声音只要 88.9 毫秒。 跑得快不代表部署门槛低。虽然每次只唤醒 30 亿参数，但 350 亿总参数还是得全装进显卡内存。光模型权重就要占掉将近 37 GiB。官方给的 vLLM-Omni 部署方案，起步就要一张 48 GB 显存的专业显卡。底层的加速算子目前也只在英伟达 Hopper 架构显卡上验证过。换到别的显卡就会退回普通模式，什么时候适配其他显卡还没时间表。日常用还有个小毛病。如果你把 [sighs] 这类动作标签或语气词放在整句话最开头，它很容易读完第一个词就一直输出静音，得把标签挪到句中才稳。另外，模型权重用的是自家的社区与研究协议，不是完全放开的 Apache 2.0 协议。对个人开发者来说，本地跑它的硬件成本太高；但对有声书、影视游戏配音和语音 Agent 团队来说，它把多角色同台、多轮讲戏和局部改词放进了一个能直接部署的开源模型里。 来源：link

A

Ariel Nurieli@wannaBfounder

16 months ago I was using you guys when you just launched, raving about how much better you are than @ElevenLabs . now you’re on the San Fran Caltrain, still better than eleven labs. Kudos @cartesia

B

Bigtega UnFiltered@BigtegaTakes

Can’t afford eleven labs? Try: Google AI Studio Speechma AI CARTESIA AI Google Vids

I

Ivan Panakhno@van_van1ch

I compared 10 voice models from @ElevenLabs, @FishAudio, and @cartesia across 312 test conversations covering 15 language and regional variants. My choice: ElevenLabs v4 Turbo. Here’s how I got there. When you listen to a generated audio clip, it’s easy to focus on how pleasant the voice sounds. In a live conversation, something else becomes very noticeable: the pause between finishing your sentence and hearing the agent respond. I focused on three things: • Voice quality and naturalness • Generation cost • TTFA, time to first audible: how long it takes to hear the start of the response To test this, I built a small evaluation setup on @livekit. An automated caller works through a scenario, the agent responds, and both sides are recorded. Everything goes into an HTML report where I can listen to conversations, compare models, inspect transcripts, and review latency measurements. The lineup included six ElevenLabs models, Fish S2 Pro and S2.1 Pro, and Cartesia Sonic 3.5 and 3.6. Languages included English, German, French, Portuguese, Japanese, Spanish, and Ukrainian, with regional variants for English, Spanish, and Portuguese. The initial scenarios had 10 turns per side. For longer comparisons, I used a scenario with 25 turns per side to hear how the voice holds up over a full conversation. Two things worth keeping in mind when running these tests: TTFA depends on the entire voice pipeline. In my setup, it measures the time from the end of the caller’s speech to the first audible agent audio frame received through LiveKit. That includes end-of-turn detection, STT, LLM, TTS, and audio delivery. It’s useful for evaluating the conversation experience, but it doesn’t isolate the speed of the TTS model. Regional wording doesn’t guarantee a regional accent. You can adapt vocabulary for British English or Spanish from Spain. That alone doesn’t guarantee the corresponding pronunciation. Accent needs a separate listening assessment. My practical takeaway: ElevenLabs v4 Turbo gives me the balance I need right now. The voice quality works for my use case, the price is acceptable, and I’m happy with the TTFA. That’s the combination I was looking for in my voice agent. What usually decides your choice of a voice model: how it sounds, response latency, or cost?

T

Tarush Agarwal@tarush_agarwal_

Your AI voice agent might take almost a full second to stop talking after a customer interrupts. We tested interruption handling across 26 text-to-speech models at Cekura. The differences were wild. Cartesia Sonic 3.5: 809ms of additional audio after a stop request. Cartesia Sonic 3.6: 764ms. Deepgram Flux: 80ms. Deepgram Aura-2: 40ms. Grok TTS: 0ms. Soniox TTS RT v2: 0ms. Some models stopped almost immediately. Others kept producing audio for nearly another second. Imagine a customer interrupting to correct their address, while the agent keeps talking over them. This is why TTS benchmarks need to measure more than voice quality and latency. Full results: link

G

Greg Helmstetter@ghelmstetter

Median time to first audio across the 8 tests: - Deepgram 0.41s - Cartesia 0.46s - OpenAI 2.05s - ElevenLabs 2.93s - Gemini 6.58s Caveat: ElevenLabs and Gemini sent the whole file at once in my setup. Both offer streaming, which starts sooner.