Sonic 3.5 is a high-performance text-to-speech (TTS) model developed by Cartesia, specifically designed for low-latency, conversational AI applications. It serves as a major update to the Sonic model family, introducing refined prosodic rhythm, natural intonation, and an expanded emotional range. The model is capable of generating realistic laughter and expressive speech that adjusts naturally to the context of the transcript, making it suitable for real-time voice agents, narration, and dubbing.
A defining technical feature of Sonic 3.5 is its State Space Model (SSM) architecture. While most modern foundation models utilize transformers, which scale quadratically with sequence length, Sonic’s SSM architecture scales linearly. This design choice enables high throughput and reduces computational costs at scale. Cartesia reports a time-to-first-audio (TTFA) latency as low as 82 to 90 milliseconds, which is intended to support seamless, human-like verbal interactions without noticeable delays.
The model natively supports 42 languages, including English, Hindi, Spanish, French, German, Japanese, and Korean. It features robust handling for alphanumerics, allowing it to read out order IDs, phone numbers, and email addresses naturally without the need for manual text normalization. Furthermore, Sonic 3.5 supports zero-shot voice cloning, which can replicate a target speaker's voice with high similarity using as little as 10 seconds of reference audio.
Capabilities and Prompting
Sonic 3.5 is designed to function with minimal prompt engineering, though it provides specific tools for fine-grained control. It includes natural alphanumeric handling where confirmation codes and serial numbers are read with appropriate pacing. When character-by-character readout is required, users can wrap text in <spell> tags. The model also allows for the adjustment of speed, volume, and emotional intensity via API parameters, enabling developers to tune the performance for specific personas.
For best results, Cartesia recommends providing well-punctuated, natural written text rather than pre-processed strings. Heteronyms such as "read," "bass," and "bow" are contextually disambiguated to ensure correct pronunciation. The model is accessible via a streaming API, allowing audio to be played back to the user as it is being generated.