Meituan LongCat-AudioDiT Revolutionizes Zero-Shot TTS Voice Cloning by Eliminating Intermediate Mel-Spectrogram Representations
The Meituan LongCat technical team has officially unveiled LongCat-AudioDiT, a pioneering model designed to push the boundaries of zero-shot Text-to-Speech (TTS) voice cloning. By fundamentally redesigning the synthesis pipeline, the model abandons traditional intermediate representations like Mel-spectrograms in favor of direct operation within the waveform latent space. Utilizing a Diffusion Transformer (DiT) architecture, LongCat-AudioDiT aims to learn the inherent laws of sound directly, thereby eliminating the cascaded errors typically associated with multi-stage data conversion. This breakthrough addresses a critical technical bottleneck in audio generation, offering a more streamlined and accurate approach to replicating human voices without the need for extensive speaker-specific training data.
