88Z AI, 1 April — The Meituan LongCat team has released the LongCat-AudioDiT model, achieving state-of-the-art performance in zero-shot voice cloning. The model performs diffusion generation directly in the waveform latent space, discarding traditional Mel-spectrogram intermediate representations to prevent information loss. It also introduces Dual Constraint Alignment and Adaptive Projection Guidance techniques to address training-inference mismatches and mitigate oversaturation.
At a glance
- Direct diffusion generation in waveform latent space avoids Mel-spectrogram information loss
- Uses DCA and APG techniques to resolve training-inference mismatches and oversaturation
The update
The Meituan LongCat team has officially launched LongCat-AudioDiT, a speech synthesis model that achieves state-of-the-art performance in zero-shot voice cloning. Unlike traditional approaches, LongCat-AudioDiT discards Mel-spectrograms as intermediate representations and performs diffusion generation directly in the waveform latent space, effectively preventing information loss. To ensure generation quality, the team introduced two key technologies: Dual Constraint Alignment (DCA) and Adaptive Projection Guidance (APG). These innovations aim to fix mismatches between training and inference stages while mitigating oversaturation in generated audio, significantly improving the naturalness and stability of speech synthesis.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 龙猫LongCatSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.