Cartesia AI

Cartesia is a voice AI company founded in 2023 by Stanford AI Lab researchers Albert Gu, Karan Goel, Chris Re, and Arjun Desai, the same team that invented the Mamba State Space Model (SSM) architecture used across machine learning. Their flagship product is Sonic, a text-to-speech API built not on transformers, but on SSMs, which process sequences by updating a compressed state vector rather than attending to every prior token. The practical result is a model that reaches 90ms time-to-first-audio (TTFA) on the standard Sonic-3 variant, and 40ms on Sonic Turbo, making it the fastest human-quality TTS API available for real-time conversational applications. Cartesia has raised approximately $122M total, with a $100M round in November 2025 from Kleiner Perkins, Index Ventures, Lightspeed, and NVIDIA.

Sonic-3, the current production model, supports 42 languages, ~130 preset voices, and fine-grained prosody controls including speed, pitch, and explicit emotion tags. Developers can inject natural laughter using a [laughter] tag directly in the text prompt. Voice cloning works from as little as 3 seconds of source audio, generating a speaker-matched voice that preserves accent, cadence, and tonal character. Beyond TTS, Cartesia's platform now includes Ink-Whisper (speech-to-text), a Voice Changer endpoint for audio transformation, an Infill endpoint for audio editing, and Line, a managed voice agent product built on top of Sonic. Enterprise deployments support on-premise and on-device inference, with SOC-2 and HIPAA compliance for regulated industries.

What Cartesia produces in April 2026

Sonic-3 is the current default model. It generates streaming audio output starting at 90ms TTFA from API call to first audio byte, measured under standard network conditions. The Turbo variant drops to 40ms at the cost of a 500-character input limit per request and a narrower quality ceiling. Both models use SSM inference, meaning audio streams token-by-token at playback speed rather than generating a complete clip and returning it. This streaming-native architecture is why voice agents built on Cartesia feel conversationally responsive: the user hears the AI begin speaking at roughly the same time a human would.

Language support as of Sonic-3 covers 42 languages with native-accent voices, up from 15 in the original Sonic and 15 in Sonic-2 at launch. Voice customization exposes speed (playback rate), pitch offset, emotion dials (calm, angry, excited, sad, and others), and the laughter injection tag. Pro Voice Cloning, available on Startup plan and above, trains a speaker model from a short audio sample and bills at 1.5 credits per character rather than the standard 1 credit per character. The platform publishes 99.9% uptime SLAs and claims the fastest P90 latencies globally across its CDN deployment.

Goodcall, a business AI phone agent company, deployed 2,217 unique voice agents on Sonic within three months of switching. Their CEO described the outcome directly:

"Sonic is the only product in existence with model latency of less than 100 ms, outperforming its next best alternative by a factor of four." - Bob Summers, CEO of Goodcall, Cartesia customer case study, 2025

Where Cartesia sits versus ElevenLabs and Resemble AI

The TTS market in 2026 is roughly divided between general-purpose content creation tools and real-time agent infrastructure. Cartesia competes on the latter axis. The mechanical differences between the three main competitors matter for choosing correctly.

ElevenLabs runs transformer-based TTS. Transformer inference attends to every previous token in the sequence, which scales well for long-form narration quality but adds overhead per token. Their Flash v2.5 model hits ~75ms TTFA, which is competitive with Cartesia Turbo. However, ElevenLabs' quality-tier models run at 300ms or more TTFA, and their voice library is substantially broader: 4,000+ preset voices across 70+ languages versus Cartesia's ~130 voices and 42 languages. In a blind preference test on 100 held-out voices, users preferred Sonic-2 over ElevenLabs Flash V2 at 61.4% to 38.6% for conversational snippets. Voice cloning for ElevenLabs requires a minimum of 10 seconds of audio; Cartesia's instant cloning works from 3 seconds. The practical split: ElevenLabs for rich voice libraries, dubbing, long-form narration, and multilingual content where language breadth matters. Cartesia for real-time agent infrastructure where every millisecond of latency is a product quality metric.

Resemble AI takes a different architectural path. Their Chatterbox Turbo model uses a 350M-parameter transformer with a distilled one-step diffusion decoder, reducing generation from 10 diffusion steps to one. Resemble's differentiation is security and authenticity: they build deepfake detection, audio watermarking, and voice forensics tooling into the platform, making it the natural choice for regulated industries that need to prove audio provenance. Chatterbox is released as open-source, which Cartesia's Sonic is not. Resemble's latency benchmarks do not match Cartesia's published TTFA numbers, and they do not lead on speed as a positioning claim. The comparison is enterprise security posture and open-source deployability (Resemble) versus raw latency performance for real-time systems (Cartesia).

"Using Cartesia's generative voice API, Sonic, we've strengthened Cresta AI Agent to move beyond rigid scripts and towards delivering empathetic, human-like conversations." - Tim Shi, Co-Founder and CTO at Cresta, Cartesia customer case study, 2025

The licensing and copyright reality

Cartesia's terms give commercial rights to audio output starting from the Pro plan ($4/mo billed annually). The Free tier restricts output to personal, non-commercial use. Pro Voice Cloning, which trains a speaker model from submitted audio, requires that users own or have rights to the audio they submit. The platform does not publish training data sources for Sonic's preset voices, and no training-data litigation has been publicly filed against Cartesia as of April 2026.

On-device and on-premise deployment options mean organizations can run Sonic inference locally, which sidesteps data retention concerns in regulated industries. Cartesia markets HIPAA compliance for healthcare voice agent deployments, and Cresta's contact center deployment at Brinks Home confirms enterprise production use. Veronica Moturi, SVP of Customer Experience at Brinks Home, noted that "Our voice AI Agent guides customers through complex, multi-step troubleshooting scenarios, and the positive feedback from these customers has been eye-opening," describing results from the Cresta deployment using Sonic.

The commercial rights structure is straightforward: pay the Pro tier or above, and generated audio is yours for commercial use. Voice clones trained from your own recordings belong to your account. Cartesia does not claim ownership or rights to output audio generated through their API.

Where Cartesia reliably falls short

The 500-character limit on Sonic Turbo is the most developer-reported frustration. Building a long-form narration or a multi-turn conversation where a single agent response exceeds 500 characters requires chunking the text, managing chunk boundaries to avoid mid-word splits, and stitching audio segments. ElevenLabs Flash v2.5 accepts up to 40,000 characters per request. For audiobook production, podcast narration, or any application where individual segments run long, this limit forces engineering work that competitors avoid.

Language coverage at 42 languages is real but still narrower than ElevenLabs' 70+. Teams building for Southeast Asian markets, less common African languages, or regional dialects with limited training data will find gaps. The preset voice library at ~130 voices is also thinner than ElevenLabs' catalog of 4,000+. Teams that want regional accent variation, age-diverse voices, or broad character coverage for gaming and media production face voice limitations that require investment in custom cloning.

Cartesia is a developer API, not a business tool. There is no native Salesforce connector, no helpdesk workflow integration, no out-of-the-box call routing. Teams describing Cartesia as their voice layer consistently note that it handles the voice well but contributes nothing to the surrounding application architecture. One common framing: Cartesia provides the engine, not the chassis, the steering, or the navigation. Developers building a complete voice agent must assemble the rest independently.

Manual emotion tuning adds overhead in dynamic conversational contexts. Unlike platforms with automatic prosody detection, Cartesia requires explicit emotion tags or manual speed/pitch dial adjustments. In a scripted system this is manageable; in a generative conversational AI responding to unpredictable user inputs, keeping emotion controls accurate requires either fine-tuned LLM prompt engineering or a classification layer that decides which emotion tag to apply.

Who Cartesia is for

Cartesia targets developers building real-time conversational AI where latency directly affects product quality. If your application is a voice agent, a phone bot, an AI avatar in a game, or any interactive system where the user is waiting for a voice response, 90ms versus 300ms TTFA is audibly different. Goodcall's 97% call interaction rate and 2,217 deployed agents represent the kind of outcome that low-latency TTS enables: users stay on the line because the AI feels like a natural conversation partner.

Healthcare and financial services teams running compliance-sensitive workloads benefit from the on-device and on-premise deployment options. HIPAA compliance with no cloud data transmission is a meaningful differentiator for that sector. Enterprises at scale who need priority support, high concurrency, and custom SLAs will find the Scale and Enterprise tiers well-suited.

Skip Cartesia when your primary needs are voice variety and language breadth for content creation. If you are producing multilingual podcasts, dubbing video content into 50+ languages, or need a library of 1,000+ voices for a creative platform, ElevenLabs' catalog will serve you better. Skip Cartesia if you need a no-code or low-code voice solution: there is no point-and-click interface for non-technical users, no campaign builder, no embedded player. The API is the product. And skip Sonic Turbo specifically if your use case generates text segments longer than a few sentences at a time, unless you are prepared to build and maintain a chunking pipeline.