stableaudioopen.app Open Source Text-to-Audio Generation for Audio Samples

Generate variable-length stereo audio samples from text descriptions using this open-source model. It produces audio at a 44.1kHz sample rate with lengths up to 47 seconds. The tool specializes in creating specific audio elements such as drum beats, instrument riffs, ambient sounds, and foley recordings, making it particularly useful for music production and sound design tasks. It is intentionally not optimized for generating complete songs, complex melodies, or human vocals.

The model utilizes a transformer-based diffusion architecture operating within the latent space of an autoencoder, conditioned by T5-based text embeddings. Training involved nearly half a million audio recordings sourced exclusively from FreeSound and the Free Music Archive, all under permissive licenses (CC0, CC BY, or CC Sampling+), ensuring no copyrighted music was included. Users can access the model weights via Hugging Face under Stability AI's non-commercial research community agreement license and utilize the associated open-source `stable-audio-tools` library for inference and fine-tuning the model on their own datasets.