Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Audio super resolution using neural networks
| Date | Stars |
|---|---|
| 2026-07-31 | 1263 |
| 2026-08-06 | 1263 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
Audio Super Resolution Using Neural Networks
============================================
This repository implements the audio super-resolution model proposed in:
```
S. Birnbaum, V. Kuleshov, Z. Enam, P. W.. Koh, and S. Ermon. Temporal FiLM: Capturing Long-Range Sequence Dependencies with Feature-Wise Modulations. NeurIPS 2019
V. Kuleshov, Z. Enam, and S. Ermon. Audio Super Resolution Using Neural Networks. ICLR 2017 (Workshop track)
```
## Installation
### Requirements
The model is implemented in Python 3.7.10 and uses several additional libraries.
* `tensorflow==2.4.1`
* `keras==2.4.0`
* `numpy==1.19.5`
* `scipy==1.6.0`
* `librosa==0.8.3`
* `h5py==2.10.0`
* `matplotlib==3.3.4`
A full list of the packages on our enviornment is in `requirements.txt`
### Setup
To install this package, simply clone the git repo:
```
git clone https://github.com/kuleshov/audio-super-res.git;
cd audio-super-res;
conda env create -f environment.yaml
conda activate audio-super-res
```
## Running the model
### Contents
The repository is structured as follows.
* `./src`: model source code
* `./data`: code to download the model data
### Retrieving data
The `./data` subfolder contains code for preparing the VCTK speech dataset.
Make sure you have enough disk space and bandwidth (the dataset is over 18G, uncompressed).
You need to type:
```
cd ./data/vctk;
make;
```
Next, you must prepare the dataset for training:
you will need to create pairs of high and low resolution sound patches (typically, about 0.5s in length).
We have included a script called `prep_vctk.py` that does that, which works as follows.
```
usage: prep_vctk.py [-h] [--file-list FILE_LIST] [--in-dir IN_DIR] [--out OUT]
[--scale SCALE] [--dimension DIMENSION] [--stride STRIDE]
[--interpolate] [--low-pass] [--batch-size BATCH_SIZE]
[--sr SR] [--sam SAM]
optional arguments:
-h, --help show this help message and exit
--file-list FILE_LIST
list of input wav files to process
--in-dir IN_DIR folder where input files are located
--out OUT path to output h5 archive
--scale SCALE scaling factor
--dimension DIMENSION
dimension of patches (use -1 for no patching)
--stride STRIDE stride when extracting patches
--interpolate interpolate low-res patches with cubic splines
--low-pass apply low-pass filter when generating low-res patches
--batch-size BATCH_SIZE
we produce # of patches that is a multiple of batch
size
--sr SR audio sampling rate
--sam SAM subsampling factor for the data (only applicable for multispeaker data)
```
The output of the data preparation step are two `.h5` archives containing, respectively, the training and validation pairs of high/low resolution sound patches.
You can also generate these by running `make` in the corresponding directory, e.g.
```
cd ./speaker1;
make;
```
This will use a set of default parameters.
To generate the files needed for the training example below, run the following from the `speaker1` directory:
```
python ../prep_vctk.py \
--file-list speaker1-train-files.txt \
--in-dir ../VCTK-Corpus/wav48/p225 \
--out vctk-speaker1-train.4.16000.8192.4096.h5 \
--scale 4 \
--sr 16000 \
--dimension 8192 \
--stride 4096 \
--interpolate \
--low-pass
python ../prep_vctk.py \
--file-list speaker1-val-files.txt \
--in-dir ../VCTK-Corpus/wav48/p225 \
--out vctk-speaker1-val.4.16000.8192.4096.h5.tmp \
--scale 4 \
--sr 16000 \
--dimension 8192 \
--stride 4096 \
--interpolate \
--low-pass
python ../prep_vctk.py \
--file-list speaker1-train-files.txt \
--in-dir ../VCTK-Corpus/wav48/p225 \
--out vctk-speaker1-train.4.16000.-1.4096.h5 \
--scale 4 \
--sr 16000 \
--dimension -1 \
--stride 4096 \
--interpolate \
--low-pass
python ../prepExcerpt of 12,026 characters
Read on GitHub
Volodymyr Kuleshov · Cornell Tech · United States
41
2
2
1
1
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:814b9c684ac13e0f, desc:super resolution