Meituan LongCat Team Open-Sources LongCat-AudioDiT Speech Synthesis Model, Zero-Shot Cloning Outperforms Industry Competitors

Meituan's LongCat team recently open-sourced the speech synthesis model LongCat-AudioDiT, offering 1B and 3.5B versions with code and weights under the MIT license. Its core feature is zero-shot voice cloning—using just a reference recording, it can read any new text in the speaker's tone.

Traditional speech synthesis converts text to a spectrogram first then to waveform, but LongCat-AudioDiT skips this step, directly generating diffusion on compressed waveform representations to reduce errors. In the Seed benchmark test, the 3.5B version achieved a Chinese speaker similarity of 0.818, surpassing ByteDance's Seed-DiT; the similarity on the difficult subset was 0.797, better than Alibaba's CosyVoice3.5.

The paper also noted a counterintuitive finding: higher restoration quality in the intermediate compression step does not necessarily lead to better final synthesis results.