What happened
Cartesia launched Sonic-3.5 and Ink-2 on June 16, presenting them as a unified real-time voice stack for building voice agents. Sonic-3.5 handles text-to-speech with a time-to-first-audio of roughly 82 to 90 milliseconds, support for 42 languages, and voice cloning. Ink-2 handles speech-to-text and ranks first for accuracy on the streaming STT leaderboard with around 100 milliseconds of latency.
Both models are built on State Space Model architecture rather than transformers, which Cartesia ties to its low-latency results. The company says the pairing makes it the only provider ranked number one on both the Artificial Analysis TTS and STT leaderboards, aimed at agents that need synthesis and recognition in one stack.
Why it matters
Latency is the main obstacle to natural voice interaction, where even small delays make conversations feel stilted. Combining leading TTS and STT in one stack simplifies the plumbing for developers building voice agents.
The use of State Space Models rather than transformers is also a notable architecture choice for real-time audio.
MintedBrain take
Leaderboard rankings and quoted latency figures are useful signals but rarely match production conditions. Test both models on your own accents, languages and network paths before committing, and watch how latency holds up under real concurrency rather than benchmark loads.
Discussion
Sign in to comment. Your account must be at least 1 day old.