AudioLDM2 is an open-source text-to-audio model published on GitHub that uses latent diffusion to generate audio, music and speech from plain-text prompts. Developed by Haohe Liu and collaborators, the project has gathered around 2,600 GitHub stars and is a reference implementation for open generative audio research.
AudioLDM2 is the successor to AudioLDM and accompanies the paper “AudioLDM 2: Learning Holistic Audio Generation with Self-supervised Pretraining”. Instead of generating waveforms directly, the model works in a compressed latent space and is conditioned on text through a “language of audio” representation, which lets one system handle sound effects, music and speech. The repository ships pretrained checkpoints, full training code and inference scripts. Released checkpoints include an improved 16 kHz model trained on more data, a 48 kHz model for high-fidelity output, and a text-to-speech checkpoint. The code was last updated in 2024, and demos and model weights are also published on Hugging Face. Because everything runs locally, it is often used as the audio backbone in research prototypes and creative tooling experiments.
AudioLDM2 is free and open source. You can download the code and checkpoints from GitHub and Hugging Face at no cost; the only expense is the GPU compute you run it on.
Audio researchers, sound designers and developers who want a proven open model for generating music, effects or speech on their own hardware.