Cartesia

Cartesia came out of the Stanford work behind state space models, and that architecture choice is the whole product thesis. Transformers are expensive to run at low latency for streaming audio. SSMs cost less per unit of audio and degrade more gracefully on small hardware, which is how Sonic reaches roughly 90ms time to first audio under good conditions and how the same model family ends up running on a laptop. The lineup is Sonic-3.6 for text to speech, Ink-2 for streaming transcription and Line for agent orchestration on top of both.

Sonic-3.6 covers 44 languages. French and German are strong. Mandarin is competent for conversational text but lacks the tonal control you want for anything performative, so a Mandarin audiobook remains an ElevenLabs job. Pricing on the subscription tiers is credit-based rather than per character, roughly one credit per character: free at 20,000 credits a month with no commercial rights, $5 for Pro at 100,000, $49 for Startup at 1.25M and $299 for Scale at 8M. Agent minutes bill separately at $0.06, which is easy to miss when you model costs.

Anyone building a real-time voice agent should put this head to head with ElevenLabs v3 Conversational, and the deciding factor is usually deployment rather than sound quality, because both clear the bar now. If audio has to stay on a device or inside a VPC, Cartesia is one of very few options that will take the call. If you want a large stock voice library, dubbing tooling or a mature integrations ecosystem, ElevenLabs is further ahead and it is not close. Worth watching whether the on-device story extends past Apple silicon, since that is what decides if it becomes useful on embedded Linux hardware in a plant.