Cartesia vs Fish Audio: pricing, features and which to choose

Cartesia is a San Francisco startup founded in 2023 by Karan Goel, Albert Gu, Arjun Desai, Brandon Yang and Chris Ré, researchers from Stanford's AI lab. Gu co-created the S4 and Mamba state space model architectures, which process long sequences with less memory than transformers, and that research is the basis of the company's products. Cartesia launched its Sonic text-to-speech model in 2024 and raised a $64 million Series A led by Kleiner Perkins in March 2025, bringing total funding to $91 million, according to Fortune. Its homepage names ServiceNow, Retell, Blue Machines and Quora among customers. The company's pitch is speed: models that start speaking quickly enough for live phone conversation, sold through one API that covers speech generation, transcription and complete voice agents.

Developers reach Cartesia through a REST and WebSocket API or the browser playground. Sonic turns text into streaming audio, with controls for speed, emotion and pronunciation, and supports more than 40 languages. Instant voice cloning creates a voice from a few seconds of audio on paid plans, and professional cloning trains a closer copy from longer recordings on the Startup plan and above. A localisation feature gives a cloned voice a new accent. Ink transcribes live audio for agents that need to hear callers. Managed Agents, the newest product, hosts complete phone agents with telephony, so a team can define an agent's prompt and tools without wiring a separate orchestration layer. Cartesia also offers on-premises and on-device deployment for customers with data residency rules.

Cartesia's voice library and creator tools are thinner than ElevenLabs, which also offers dubbing, sound effects and music, so it is a weaker fit for video producers and audiobook narrators. Managed Agents are newer than Vapi or Retell and have fewer integrations. Credit-based billing means long-form narration gets expensive faster on the low tiers. ElevenLabs Flash competes directly on latency with a larger voice catalog. Deepgram's Aura voices are cheaper per minute but less expressive, and Hume, once a rival on expressive speech, is closing its TTS API in November 2026. Pick Cartesia when you are building a real-time voice agent or interactive app and the delay before the first syllable is the metric you care about, or when you need on-device or on-premises speech models.

Fish Audio began as Fish Speech, a project by former Nvidia researcher Shijia Liao, who trained a voice-generation model on one GPU and open-sourced it. The company later rebranded as Fish Audio, with Rissa Cao serving as CEO and co-founder. Its public history records the first Fish Speech release in December 2023. The company launched five models in the year before July 2026, including four speech-generation systems and one speech-to-text model. TechCrunch reported that Fish Audio raised a $52 million seed round in July 2026, led by Coreline Ventures and Capital Today. The report also cited more than eight million users across hosted and open-source models, $21 million in annual recurring revenue, and a GitHub repository exceeding 31,000 stars.

A typical Fish Audio workflow begins by selecting a voice or creating one through cloning or Voice Design, then entering a script and adding bracketed directions such as [whisper] or [laugh]. Users generate audio, review it, and download or share the result. For production integrations, developers send text to the REST TTS endpoint or use WebSocket streaming through Python or TypeScript SDKs. S2.1-Pro supports multi-speaker dialogue and 83 languages; s2.1-pro-free suits development under fair-use limits. For transcription, teams upload a recording to /v1/asr and specify the transcribe-1-pro header to receive speaker markers, turns, timestamps, and vocal-event cues. Live applications instead keep a realtime WebSocket open with transcribe-stream-1, billed by session duration.

Fish Audio differs from ElevenLabs, Cartesia, and PlayHT in its combination of expressive controls, open models, and deployment options. S2.1 Pro offers 83 languages, multi-speaker dialogue, and free development access, while bracket-based direction and open-weight models provide control over delivery and deployment. The $15-per-million-UTF-8-byte API price differs from credit-based web plans, complicating budgeting across products. Starter API accounts allow five concurrent requests, and long recordings can occupy a slot for minutes. Its terms permit use of Usage Data and Content to develop or enhance models, while voice submissions and public content carry licensing and consent responsibilities. Fish Audio suits creators and engineering teams prioritizing expressive control, multilingual output, and self-hosting.