Sonic: A Low-Latency Voice Model for Realistic Speech 198

Sonic, a low-latency real-time inference voice generation model based on state space models, is a real-time voice generation model developed by Cartesia ai. Built on their proprietary state space models, it is designed for real-time interactive voice applications.

Notable features of Sonic include:

Technical Foundation:

• Sonic is built on the next-generation state space model, an advanced deep learning architecture.

• The model achieves the fastest text-to-speech conversion on the market, with end-to-end latency under 200 milliseconds, and a model latency of just 135 milliseconds—the fastest among comparable models.

Performance Characteristics:

• Extreme speed: Sonic's latency is only 135 milliseconds, ensuring real-time response, which is critical for interactive applications.

• High throughput: Leveraging the first-of-its-kind state space model inference stack, Sonic supports high concurrency and low-cost inference, making it suitable for large-scale deployment.

Voice Quality:

• Ultra-realistic voice: Sonic is capable of generating emotional and expressive human-like speech, significantly enhancing the naturalness and realism of voice synthesis.

Personalization Features:

• Zero-shot voice cloning: With just 10 seconds of recording, Sonic can match tone, intonation, and vocal characteristics, enabling personalized voice cloning.

• Controllable parameters: Users can adjust parameters such as pitch, speech rate, and emotion to achieve personalized voice design.

Cartesia, a startup focused on building real-time intelligence, delivers high-quality real-time voice experiences by providing efficient, long-lifecycle real-time intelligence for every device through its innovative State Space Model (SSM) technology.

Details: https://cartesia.ai/blog/sonic