Speech-02-Turbo is a low-latency text-to-speech (TTS) model developed by MiniMax, a Shanghai-based AI company. Released in April 2025 as part of the Speech-02 series, it is specifically optimized for real-time interactive applications that require rapid response times without compromising vocal quality.
The model supports synthesis and zero-shot voice cloning for over 30 languages, including English, Mandarin, Cantonese, Japanese, and Korean. It provides access to an expansive library of more than 300 pre-built voices and allows users to clone specific timbres using as little as 10 seconds of reference audio. MiniMax reports that this process achieves high vocal similarity while maintaining natural native accents across diverse linguistic contexts.
Technically, Speech-02-Turbo is built on an autoregressive Transformer architecture. It integrates a learnable speaker encoder for extracting timbre features and a Flow-VAE (Variational Autoencoder) to enhance overall audio fidelity and consistency. The model includes granular controls for emotional expression—supporting tones such as happy, sad, and urgent—and offers adjustable parameters for pitch, speed, and volume.