AudioSR

AudioSR performs audio super-resolution: it takes recordings sampled at low rates such as 4kHz, 8kHz, 16kHz, or 24kHz and reconstructs them at 48kHz, restoring missing high-frequency content with a diffusion-based model. It is deliberately versatile, handling speech, music, and environmental or animal sounds rather than a single domain, and was published at IEEE ICASSP 2024 by Haohe Liu and collaborators. Two checkpoints are provided, a general basic model and a speech-specialized one, and inference exposes DDIM step count, guidance scale, and a seed for reproducible results. The tool installs from PyPI as a pip package, processes single files or batch lists from the command line, and includes a Gradio web demo, with hosted inference also available on Replicate. One documented limitation: it works well on audio with clean low-pass cutoff patterns but can struggle with artifacts like MP3 compression, so the authors recommend low-pass filtering problematic input first. A GPU is recommended for practical processing speeds.