UniverSR (Super Resolution)
UniverSR - audio super resolution algorithm for restoring high frequencies. It's based on paper UniverSR: Unified and Versatile Audio Super-Resolution via Vocoder-Free Flow Matching (ICASSP 2026).
Supports 8/12/16/24 kHz input to 48 kHz output. Two models available: universal (all types of audio) and speech.
Auto input bandwidth mode: if the input file is 8/12/16/24 kHz, its native sample rate is used; for 44.1/48 kHz files the effective bandwidth is assumed to be 16 kHz (typical for old recordings and lossy codecs), everything above 8 kHz gets restored; other sample rates (11.025/22.05/32 kHz) are rounded down to the closest supported rate. You can also set the input bandwidth manually (8000/12000/16000/24000) - it should not be higher than the real bandwidth of the recording, otherwise real high frequencies will be replaced by generated ones.
Guidance scale controls the density of generated high frequencies: higher values give brighter but less accurate restoration (at 3.0+ it causes hiss and clipping). Recommended values: 1.0-1.5 for speech, 1.5-2.0 for music. Default: 1.5.
Original repository: https://github.com/woongzip1/UniverSR
Pretrained weights: woongzip1/universr-audio, woongzip1/universr-speech
