Meituan LongCat
LongCat-AudioDiT is an open-weight diffusion-based text-to-speech model from the Meituan LongCat team. The repository and paper describe a non-autoregressive TTS system that operates directly in waveform latent space instead of mel-spectrograms, reducing pipeline complexity and compounding errors. Its inference path combines a waveform VAE with a diffusion backbone, corrects a training-inference mismatch and uses adaptive projection guidance rather than traditional classifier-free guidance. The repository releases code and model weights for research, provides Hugging Face-compatible implementation and inference scripts, and reports state-of-the-art zero-shot voice cloning results on the Seed benchmark. The largest LongCat-AudioDiT-3.5B model is reported to improve speaker similarity on Seed-ZH and Seed-Hard over the previous Seed-TTS baseline.
Editorial verdict
Researchers and speech teams evaluating open-source TTS, waveform-latent diffusion and zero-shot voice cloning.
Avoid treating it as a turnkey production voice platform until runtime, rights and deployment constraints are validated.
LongCat-AudioDiT belongs in AI Audio because it is a direct-text-to-speech and voice-cloning model with released code and weights, not a generic research paper.
Open-source MIT repository and released model weights; inference runs locally or through a Hugging Face-compatible workflow
GitHub repository, Model weights download, Local inference, Hugging Face-compatible workflow
MIT covers the repo, but voice cloning rights, model weights and generated-audio use still need explicit review.
Prompt audio handling, retained voice samples and generated audio storage should be reviewed before production use.
Evaluate when you need prompt audio plus text to reproduce speaker style and voice similarity.
Useful for comparing against Seed-TTS, CosyVoice, Qwen3-TTS and MiniMax speech baselines.
The model and scripts can be used to study waveform-latent diffusion inference and guidance methods.
minimax-audio
stepaudio
qwen-audio
zhipu-glm-audio
official · en · verified 2026-08-04
Confirms repository name, MIT license, model description, seed benchmark results, code and weights release, and inference usage.
docs · en · verified 2026-05-18
Source paper for waveform-latent diffusion, APG guidance and test claims.
Last checked: 2026-08-04