Skip to main content
Cartesia Sonic is a real-time speech synthesis model built around streaming from the ground up. The current generation, Sonic-3.6, is what Cartesia describes as its most natural and human-sounding model, and the company reports it at number one on Artificial Analysis’s Speech Arena leaderboard.
Available in Ollang Workflows as cartesia — AI Dubbing. See the Text-to-Speech catalog.

​
Model versions

Earlier generations — sonic-3, sonic-3.5, sonic-2, sonic-turbo, and sonic — remain documented by Cartesia.

​
Coverage

44 languages, with Odia and Urdu added in 3.6 and Hinglish explicitly supported: en de es fr ja pt zh hi ko it nl pl ru sv tr tl bg ro ar cs el fi hr ms sk da ta uk hu no vi bn th he ka id te gu kn ml mr pa or ur That is 15 more than ElevenLabs Multilingual V2, and it includes South Asian languages — Telugu, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Odia — that most premium engines do not serve.

​
Key capabilities

  • Automatic emotional calibration — Sonic reads the emotional subtext of the transcript and adjusts delivery without explicit direction, so an undirected line still lands with appropriate affect.
  • Inline non-verbal expression — laughter, sighs, and disfluencies are written directly into the transcript rather than requiring separate markup or a second pass.
  • Contextual pacing, correct alphanumeric read-out, and heteronym disambiguation — the three places synthesized narration most often gives itself away.
  • Instant voice cloning from roughly 10 seconds of reference audio, which keeps speaker similarity high enough for brand consistency across a catalog.
  • Streaming-first architecture with sub-90 ms stated latency, which translates into fast batch throughput for dubbing volume.

​
Where it fits in a workflow

Reach for Cartesia on conversational and dialogue-heavy content, on high-volume dubbing where turnaround matters, and whenever your language list runs past ElevenLabs’ 29 without needing the full 99 that Gemini TTS reaches.

​
Trade-offs

  • Newer model line, so there is less accumulated production history behind it than ElevenLabs at studio-narration quality.
  • Published latency is model latency, not an end-to-end round trip. Benchmark against your own pipeline if latency is a hard requirement.
The in-product description for this provider still references an earlier Sonic generation with 15 languages. Sonic 3.6 supports 44.

​
Reference