A two-stage U-Net for high-fidelity denoising of historical recordings
Official repository of the paper (not submitted yet):
E. Moliner and V. Välimäki,, "A two-stage U-Net for high-fidelity denosing of historical recordinds", in Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Singapore, May, 2022
Abstract
Enhancing the sound quality of historical music recordings is a long-standing problem. This paper presents a novel denoising method based on a fully-convolutional deep neural network. A two-stage U-Net model architecture is designed to model and suppress the degradations with high fidelity. The method processes the time-frequency representation of audio, and is trained using realistic noisy data to jointly remove hiss, clicks, thumps, and other common additive disturbances from old analog discs. The proposed model outperforms previous methods in both objective and subjective metrics. The results of a formal blind listening test show that the method can denoise real gramophone recordings with an excellent quality. This study shows the importance of realistic training data and the power of deep learning in audio restoration.
Listen to our audio samples
Requirements
You will need at least python 3.7 and CUDA 10.1 if you want to use GPU. See requirements.txt for the required package versions.
To install the environment through anaconda, follow the instructions:
conda env update -f environment.yml
conda activate historical_denoiser
Denoising Recordings
Run the following commands to clone the repository and install the pretrained weights of the two-stage U-Net model:
git clone https://github.com/eloimoliner/denoising-historical-recordings.git
cd denoising-historical-recordings
wget https://github.com/eloimoliner/denoising-historical-recordings/releases/download/v0.0/checkpoint.zip
unzip checkpoint.zip /experiments/trained_model/
If the environment is installed correctly, you can denoise an audio file by running:
bash inference.sh "file name"
A ".wav" file with the denoised version, as well as the residual noise and the original signal in "mono", will be generated in the same directory as the input file.
Training
TODO
49 Oct 31, 2022
21 Oct 06, 2022
2 Oct 25, 2022
72 Jan 02, 2023
551 Dec 31, 2022
12 Apr 28, 2022
15 Sep 26, 2022
2 Aug 20, 2022
1 Jan 16, 2022
123 Dec 13, 2022
22 Nov 25, 2022
173 Dec 25, 2022
20 Nov 15, 2022
733 Dec 29, 2022
1.2k Dec 21, 2022
272 Dec 28, 2022
134 Dec 30, 2022
6 Dec 06, 2022
1 Dec 31, 2021
103 Dec 31, 2022