Cartesia Sonic-3

Cartesia Sonic-3 is an AI narration tool. Cartesia's Sonic-3 is a streaming text to speech API built for low-latency voice agents and expressive narration. It offers sub-90ms first audio, 44 languages, instant voice cloning and custom pronunciation dictionaries, with free, self-serve paid and enterprise deployment tiers.

Sonic-3, now offered as Sonic-3.6, is Cartesia's text to speech model delivered as a streaming API. The headline claim is first audio in under 90 milliseconds, with playback starting before the full response has been generated. That latency profile is the reason the product is framed around live conversations rather than pre-rendered narration, though nothing stops a team from using it for recorded content. Expression is handled at the transcript level. Sonic reads the emotional subtext of the input and calibrates delivery on its own, and non-verbal cues such as laughter can be typed directly into the script. Voice cloning takes about 10 seconds of reference audio for an instant clone, with a higher-fidelity professional tier built from 15 to 30 minutes of clean speech and custom voice development available under enterprise contracts. Pronunciation dictionaries let teams pin down proper nouns, order identifiers and domain jargon, and localization carries tone and speaker identity across 44 languages. The model sits on State Space Model architecture developed by Cartesia's founding team at Stanford. The vendor's argument is that this avoids the usual trade between quality and speed, and it points to first place on Artificial Analysis blind leaderboards for both speech and speech to text. The practical audience is builders: voice agent platforms, support automation, recruiting and sales outreach, plus media teams that need expressive narration. Deployment ranges from self-serve API keys to VPC, on-premise and air-gapped installs, backed by SOC 2 Type 2, HIPAA eligibility, GDPR compliance and optional zero data retention. Competitively, Sonic is aimed at the hosted speech APIs from the large labs and at specialist voice vendors. Cartesia's differentiators are latency, deployment flexibility for regulated buyers, and a pronunciation engine that handles alphanumerics. Its weaknesses are the ones that come with an API-first product: there is no timeline editor for post-production work, and the compliance features that matter most to enterprises sit behind sales conversations.