Cartesia

Cartesia is an artificial-intelligence company building real-time voice models designed for streaming, on-device and low-latency use. Its products are Sonic, a text-to-speech model, Ink, a speech-to-text model, and Line, a platform for building and running voice agents on top of both.^[1][2]

History

Cartesia was founded in September 2023 by Stanford AI Lab researchers Karan Goel, Albert Gu, Arjun Desai and Brandon Yang, together with Stanford professor Christopher Ré, whose group invented the state space model family of architectures — notably S4 — as a more efficient alternative to transformers for long, continuous sequences such as audio.^[3] The company was formed specifically to productise that architecture.

Its first product, Sonic, launched as a streaming text-to-speech model able to produce its first audio byte in around 90 milliseconds. Early funding of $22 million led by Index Ventures brought total capital to $27 million; a $64 million Series A led by Kleiner Perkins in March 2025 took the total to about $91 million and shipped Sonic 2.0, twice the size of its predecessor while running faster — 90 ms for the full model and 40 ms for the turbo variant — with blind preference tests favouring it over the next best provider.^[4][5] A $100 million Series B in late 2025, with NVIDIA joining as a strategic investor, accompanied the release of Sonic-3, supporting around 40 languages at roughly 190 ms latency. Sonic 3.6 followed in August 2026 with 44 languages.^[2]

Key Concepts and Technology

The defining choice is the state space model rather than a transformer. SSMs maintain a fixed-size hidden state that is updated as a sequence streams past them, giving constant-time-per-token behaviour and memory use that does not grow with sequence length — which is what makes sub-100-millisecond speech generation and on-device deployment practical.^[3]

Around that core, Cartesia has built developer infrastructure: streaming APIs, voice cloning, a voice-changer endpoint, and infill editing for correcting or extending generated audio, delivered with enterprise controls (SOC 2 and HIPAA compliance, on-premise and on-device options) and a serving stack targeting high uptime and low P90 latency.^[4] Evaluation is done through third-party blind human-preference tests rather than internal scores, a norm the voice-model market adopted partly because intelligibility and naturalness resist single-metric scoring.

Applications and Impact

Voice agents are the primary commercial surface: customer-service lines, appointment scheduling, inbound qualification, and in-product voice assistance, where a pause of more than a few hundred milliseconds reads as a broken system. Other uses include dubbing and localisation, accessibility narration, e-learning content, game and media prototyping, and rapid audio localisation — generating the same script in many languages and accents without re-recording. Cartesia competes with ElevenLabs, OpenAI's and Google's voice stacks, and a long tail of API providers, and is one of the few independents to have raised at scale against those laboratories.

>See Also

• Voice Agents
• Text-to-Speech
• Speech Recognition
• Mamba-SSM
• Voice Cloning
• ElevenLabs
• PDPA Malaysia