fyremael / cqtdiff

Official repository of the paper "Solving Audio Inverse Problems with a Diffusion Model", submitted to ICASSP 23

CQTDiff: Solving audio inverse problems with a diffusion model

Official repository of the paper:

E. Moliner,J. Lehtinen and V. Välimäki, "Solving audio inverse problems with a diffusion model", submitted to IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Rhodes, Greece May, 2023

Read the paper in arXiv Listen to our audio samples

Setup

This repository requires Python 3.8+ and Pytorch 1.10+. Other packages are listed in requirements.txt.

To install the requirements in your environment:

pip install -r requirements.txt

To install the pre-trained weights and download a set of audio samples from the MAESTRO test set, run:

bash download_weights_and_examples.sh

Training

To retrain the model, run:

mkdir experiments/my_experiment
python train.py  model_dir="experiments/my_experiment"

To change the configuration, override the hydra parameters (listed in conf/conf.yaml)

By default, the training scripts log to wandb. Set log=False if this is not desired.

python train.py log=False

Testing

To easily test our method, we recommend running the Colab Notebook, where some of the experiments are implemented.

To run it locally, use:

python sample.py \
        inference.load.load_mode="from_directory" \
        inference.load.data_directory="$path_to_audio_files" \
        inference.mode=$test_mode

The variable $test_mode selects the type of experiments. Examples are: "bandwidth_extension", "inpainting" or "declipping". There are many other parameters to select listed in the inference section from conf/conf.yaml. Some experiment examples are located in the directory scripts/.

Remarks

The model is trained using the MAESTRO dataset, the performance is expected to decrease in out-of-distribution data.