Sonic 3.6

Sonic 3.6 is a streaming text-to-speech (TTS) model developed by Cartesia, engineered for high-fidelity audio generation with extremely low latency. It is built using a State Space Model (SSM) architecture, which provides significant efficiency gains over traditional Transformer-based systems. This architecture allows the model to begin streaming audio in under 90 milliseconds, making it a primary choice for real-time applications such as conversational AI agents, interactive gaming, and live localized broadcasts.

The model supports over 44 languages and is designed to handle complex linguistic tasks, including native code-switching and the accurate pronunciation of alphanumeric sequences. Sonic 3.6 incorporates advanced prosody and rhythm controls, enabling the generation of expressive, lifelike speech that maintains natural intonation across various emotional contexts. It is frequently cited as a top-performing model on industry benchmarks for voice naturalness and controlled synthesis.

Guidance and Controls

Sonic 3.6 provides granular control over the output through several configuration parameters. Users can modulate the speed of generation (ranging from 0.6 to 1.5), adjust the volume (0.5 to 2.0), and specify emotion settings to suit the desired persona. The model also features instant voice cloning capabilities, requiring as little as 10 seconds of clear reference audio to replicate a specific voice's characteristics and accent.

To achieve optimal results, transcripts should be well-punctuated, as the model utilizes punctuation to determine natural pauses and emphasis. For specialized vocabulary or proper nouns, Sonic 3.6 supports the International Phonetic Alphabet (IPA) for precise pronunciation overrides. Additionally, expressive markers such as [laughter] can be included directly in the text transcript to trigger non-verbal vocalizations and enhance the realism of the output.