woongzip1/universr-speech

UniverSR - Speech Only

Vocoder-free speech super-resolution model that upsamples 8/12/16/24 kHz → 48 kHz using flow matching in the complex STFT domain. Trained on speech data for VCTK benchmark evaluation.

For general use across speech, music, and sound effects, see universr-audio (recommended).

Paper: arXiv:2510.00771 Demo: woongzip1.github.io/universr-demo | Code: github.com/woongzip1/UniverSR

Usage

import torchaudio
from universr import UniverSR

model = UniverSR.from_pretrained("woongzip/universr-speech", device="cuda")
output = model.enhance("low_res_speech.wav", input_sr=16000)
torchaudio.save("output_48k.wav", output.cpu(), 48000)

Citation

@inproceedings{choi2026universr,
  title     = {{UniverSR}: Unified and Versatile Audio Super-Resolution via Vocoder-Free Flow Matching},
  author    = {Choi, Woongjib and Lee, Sangmin and Lim, Hyungseob and Kang, Hong-Goo},
  booktitle = {IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)},
  year      = {2026}
}