Arjun Desai is a cofounder of Cartesia, building real-time multimodal intelligence for conversations that demand immediate, expressive responses. His path into voice AI began with a different challenge: making medical-image reconstruction reliable under the messy conditions of clinical practice.
Desai studied biomedical engineering and computer science at Duke University before pursuing doctoral research in electrical engineering at Stanford, where he worked with Akshay Chaudhari and Christopher Ré on machine-learning systems for magnetic resonance imaging. He was the first author of VORTEX, which uses physics-informed augmentation and consistency training to improve accelerated MRI reconstruction; the work received the best-paper award at Medical Imaging with Deep Learning 2022. He also led SKM-TEA, a knee-MRI dataset and evaluation framework focused on clinically meaningful results, and developed meddlr, an open-source framework for medical-image reconstruction and analysis.
He subsequently cofounded Cartesia alongside Karan Goel, Albert Gu, Brandon Yang, and Ré, applying efficient model architectures to interactive speech. Its products include Sonic-3.6 and Ink-2, models for speech generation and recognition.
- Voice latency is a system-wide budget. Speech recognition, language-model reasoning, and synthesis share the brief window before a conversation starts feeling unnatural. Faster speech generation gives the reasoning model more time, while interruptions, background noise, and uneven response times determine the experience users actually perceive.
- State-space models enable responsive inference. By maintaining recurrent state during generation, these architectures can avoid repeatedly processing expanding histories. Desai treats low latency and expressive, high-quality speech as simultaneous requirements for practical voice agents.
- Controllability matters as much as naturalness. Useful synthetic speech adapts its accent, pacing, pronunciation, and acoustic texture to its setting. Desai also supports licensed voice marketplaces that let performers make their distinctive voices available for commercial applications.
- Orchestrated voice agents remain easier to control. Separate recognition, reasoning, and synthesis components make production systems easier to inspect and tune than emerging end-to-end speech models, particularly in sensitive applications such as healthcare.
- On-device voice inference can reduce conversational delay. Local models eliminate network round trips when cloud latency dominates, extending interactive systems beyond centralized infrastructure.
Desai’s longer-term ambition is interactive intelligence that responds across multiple modalities and operates wherever people need it.
