Speech-02-HD is a high-definition text-to-speech (TTS) model developed by MiniMax for professional-grade audio synthesis. As the high-fidelity flagship of the Speech-02 series, it is optimized for clarity and expressiveness in applications such as narration, audiobooks, and commercial voiceovers. The model supports over 32 languages and provides native-level pronunciation across diverse regional accents.
The model's architecture is built on an autoregressive Transformer integrated with a learnable speaker encoder. This configuration facilitates "Intrinsic Zero-Shot" text-to-speech, allowing the model to extract and replicate timbre features from reference audio without requiring transcriptions. Additionally, a Flow-VAE (Variational Autoencoder) is employed to improve the overall naturalness and acoustic quality of the synthesized speech.
Key capabilities of Speech-02-HD include zero-shot and one-shot voice cloning, as well as granular control over emotional tone (such as happiness, sadness, or anger) and speech attributes like pitch, speed, and volume. The system is designed to handle long-form content, supporting inputs up to 200,000 characters for asynchronous processing while maintaining consistent vocal performance throughout the output.