AudioLDM-S-Full-v2

AudioLDM-S-Full-v2 is the full-size 'small' checkpoint of AudioLDM, a text-to-audio system from the Centre for Vision, Speech and Signal Processing at the University of Surrey. Presented at ICML 2023, AudioLDM learns continuous audio representations from CLAP latents in a latent diffusion framework, generating sound effects, human speech and music from natural-language prompts.\n\nThe S-Full-v2 weights are the standard reference checkpoint loaded by the diffusers AudioLDMPipeline, supporting controllable audio length and quality via inference steps, plus zero-shot text-guided audio manipulations such as style transfer.\n\nBecause it was trainable on AudioCaps with a single GPU while reaching state-of-the-art TTA quality at the time, it remains one of the most accessible open text-to-audio models to run locally.