Cartesia Sonic

Acerca de Cartesia AI

Fundado en 2023 · San Francisco, California, USA

Cartesia AI was founded in 2023 by Karan Goel and Albert Gu, the academic team behind the influential state-space model line of research at Stanford and CMU (S4, H3, Mamba, Mamba-2). Co-founders also include Arjun Desai, Brandon Yang and Chris Re (Stanford advisor). The company set out to build voice and multimodal foundation models on a state-space backbone instead of Transformers in order to achieve sub-100 ms latency and stream-friendly inference. Cartesia raised a $27M seed in March 2024 led by Index Ventures with participation from Conviction, A* and Lightspeed, followed by a $64M Series A in October 2024 led by Kleiner Perkins at a reported $325M valuation. Sonic, the company's first product, launched in May 2024 and quickly became one of the lowest-latency commercial TTS systems on the market.

Arquitectura

State-space model (Mamba family) text-to-speech with neural codec output

Cartesia Sonic is a streaming text-to-speech model built on the structured state-space modelling (SSM) architecture pioneered by Cartesia's founders (S4, H3, Mamba, Mamba-2). Unlike Transformer-based TTS systems that scale quadratically with sequence length, Sonic uses linear-time SSMs with selective scan, which lets the model maintain a small constant-memory recurrent state and generate audio chunks as text streams in. Cartesia reports a model first-byte latency of around 75-90 ms on their hosted API, which is faster than ElevenLabs Turbo and OpenAI TTS-1. Sonic outputs 24 kHz PCM via a neural codec decoder, supports streaming text input (so it can start speaking before the LLM finishes its sentence) and offers voice cloning from short reference samples (3-30 seconds). Sonic 2 added improved prosody, multilingual coverage (15+ languages) and reduced WER. The system is offered exclusively as a hosted API; weights are not released.

Parámetros

Undisclosed

Contexto

24,000 tokens

Capacidades

• Sub-100 ms first-byte latency suitable for real-time voice agents
• State-space (Mamba-family) backbone with linear-time inference
• Streaming text input and streaming audio output via WebSocket or gRPC
• Instant voice cloning from short reference audio
• Multilingual: English, Spanish, French, German, Portuguese, Mandarin, Japanese and more
• Emotion and pace controls via inline tags
• 24 kHz PCM, MP3 and Opus output formats
• Best for: real-time voice agents, IVR systems, conversational AI, low-latency phone bots

Entrenamiento y licencia

Cartesia has not disclosed the training corpus. Public statements describe a 'diverse multilingual speech dataset' with permissioned voice talent for the stock voice library.

Licencia: Proprietary commercial API. Voice clones produced from customer audio remain customer property under the Terms of Service; commercial use is permitted on paid tiers.

Pruebas de seguridad: Voice-cloning policy requires explicit consent statements; outputs include an inaudible watermark and audio fingerprinting to detect misuse.