Deep Audio Waveform Prior (Interspeech 2022)

Official PyTorch implementation of the paper: "Deep Audio Waveform Prior".

Abstract

Convolutional neural networks contain strong priors for generating natural looking images [1]. These priors enable image denoising, super resolution, and inpainting in an unsupervised manner. Previous attempts to demonstrate similar ideas in audio, namely deep audio priors, (i) use hand picked architectures such as harmonic convolutions, (ii) only work with spectrogram input, and (iii) have been used mostly for eliminating Gaussian noise [2]. In this work we show that existing SOTA architectures for audio source separation contain deep priors even when working with the raw waveform. Deep priors can be discovered by training a neural network to generate a single corrupted signal when given white noise as input. A network with relevant deep priors is likely to generate a cleaner version of the signal before converging on the corrupted signal. We demonstrate this restoration effect with several corruptions: background noise, reverberations, and a gap in the signal (audio inpainting).

Getting Started

Prerequisites

• Linux or macOS
• Python>=3.7
• NVIDIA GPU + CUDA CuDNN (Not mandatory bur recommended)

Installation

1. Clone the repo:

   git clone https://github.com/Arnontu/DeepAudioWaveformPrior.git
   cd DeepAudioWaveformPrior

2. Install dependencies
   • torch
   • torchvision
   • torchaudio
   • librosa
   • matplotlib
   • pesq
   • julius
   • tqdm
   • scipy

Preparing your Data

To evaluate denoising (clean,noisy) audio pairs are required, but you can provide only clean samples. The source to should be specified using --source and it's format explained in the following

1. You want to provide only clean samples Matching noisy will be generated accoridngly to the given arguments. Source optional formats:
   1. Directory: will consider all supported audio files (also in sub directories)
   2. File: excpect each line to be a path to an clean audio

   /path/to/clean_example_1.wav /path/to/clean_example_2.wav

2. You want to provide (clean,noisy) Source expected to be a File. Each line contains a clean path and a noisy path, separated by a comma with the clean path first.

/path/to/clean_example_1.wav,/path/to/noisy_example_1.wav

Running

#!./scripts/denoise.sh
SOURCE=#<path/to/source>
python3.7 denoise.py --experiment_repetition 20 --show_every 250 --noise_class GAUSSIAN --noise_std 0.2 --snr 2.5 --depth 2 --lstm_layers 4 --source $SOURCE --samplerate 16000 --clip_length 2

Acknowledgments

Our experiments based on the Demucs(v2) model

Contact

For any inquiry please contact by email: arnon.turetzky@mail.huji.ac.il

Citation

If you find this useful for your research, please use the following:

@inproceedings{turetzky22_interspeech,
  author={Arnon Turetzky and Tzvi Michelson and Yossi Adi and Shmuel Peleg},
  title={{Deep Audio Waveform Prior}},
  year=2022,
  booktitle={Proc. Interspeech 2022},
  pages={2938--2942},
  doi={10.21437/Interspeech.2022-10735}
}