Cartesia Sonic-3Real-time, multilingual text-to-speech with sub‑90 ms latency and voice cloning

Cartesia Sonic-3 is a text‑to‑speech model that focuses on delivering natural, expressive speech in real time. It targets enterprises and developers who need high‑quality audio for customer‑facing applications such as marketing calls, sales outreach, and automated support. The model is built on state‑space architecture, which the vendor claims provides sub‑90 ms latency while maintaining a ranking of #1 for naturalness. The service supports more than 40 languages and a variety of regional accents, allowing a single voice model to be used across global markets. Voice cloning is offered with as little as ten seconds of source audio, producing a synthetic voice that retains the speaker’s identity. Users can also upload custom pronunciation dictionaries to ensure proper rendering of domain‑specific terms, proper nouns, or brand names. Sonic‑3’s expressive capabilities include the ability to convey emotion, tone, and even laughter, aiming to preserve the nuance of the original speaker during localization. The platform is positioned as an enterprise‑grade solution, with compliance certifications such as HIPAA, SOC 2 Type 2, GDPR, and PCI, and options for both cloud and on‑premise deployment. Typical workflows involve integrating the API into marketing automation, CRM, or contact‑center platforms to generate personalized audio messages at scale. The vendor highlights use cases like warm‑lead outreach, real‑time sales objection handling, automated customer authentication, and lifecycle‑stage follow‑ups. Security and compliance are emphasized for regulated industries. Limitations noted in the public material include the need to contact sales for pricing and onboarding, and the reliance on a cloud or managed deployment model for most customers. While the language coverage is broad, it is limited to the 40+ languages explicitly supported, and the voice cloning feature may not capture the full expressive range of a speaker with only ten seconds of audio.