Mini toolsLongCat AudioDiT 3.5B
By Meituan
Model family: LongCat
LongCat-AudioDiT-3.5B is a non-autoregressive diffusion TTS model that generates speech directly in waveform latent space using a Wav-VAE and diffusion backbone. It is built for high-quality text-to-speech and voice cloning with a simpler pipeline than multi-stage acoustic approaches. The model card says it improves on prior Seed benchmark results, reaching 0.818 SIM on Seed-ZH, 0.786 EN SIM, and 0.797 on Seed-Hard, while remaining competitive on intelligibility. It supports standard TTS and prompt-audio voice cloning, and is released under the MIT license.
Audio
Overview
LongCat-AudioDiT-3.5B is Meituan LongCat’s diffusion-based text-to-speech model built directly in waveform latent space rather than mel-spectrogram space. It is designed for high-fidelity speech generation and zero-shot voice cloning, supports Chinese and English, and is positioned as a top-performing open model on the Seed benchmark for speaker similarity and intelligibility.
About Meituan
Chinese local services platform connecting consumers with food delivery, restaurant booking, travel, retail, grocery, and on-demand neighborhood services.
Industry:
Internet Marketplace Platforms
Company Size:
111298
Location:
Beijing, Chaoyang District, CN
Website:
meituan.com
Last updated: July 9, 2026
