LongCat-AudioDiT-1B: A High-Fidelity Waveform Latent Diffusion Text-to-Speech Model
meituan-longcat/LongCat-AudioDiT-1B
AI BriefingThe diffusion-based TTS model open-sourced by Meituan's LongCat team eliminates the traditional mel-spectrogram intermediate representation and operates directly in waveform latent space, synthesizing speech using only a Wav-VAE and a diffusion backbone. The model fixes the training–inference mismatch and replaces classifier-free guidance with adaptive projection guidance. The largest version, at 3.5B parameters, achieves SOTA zero-shot voice cloning on the Seed benchmark, with speaker similarity (SIM) scores of 0.818 on Seed-ZH and 0.797 on Seed-Hard, surpassing the previous best, Seed-TTS. The research also finds that Wav-VAE reconstruction fidelity does not positively correlate with final synthesis quality.
Source: Meituan LongCat: New HuggingFace model · huggingface.co