A utility for wrapping the Free Spoken Digit Dataset into PyTorch-ready data set splits.

The Free Spoken Digit Dataset is an open data set consisting of audio recordings of various individuals speaking the digits from 0-9, with 50 recordings of each digit per individual.

The data set can be thought of as an audio version of the popular MNIST data set which consists of hand-written digits. However, the fact that the data consists of recordings of different durations makes it more challenging to deal with than the fixed-size images of MNIST.

TorchFSDD aims to provide an interface to FSDD for PyTorch model development, by providing a torch.utils.data.Dataset wrapper that is ready to be used with a torch.utils.data.DataLoader.

master	dev

from torchfsdd import TorchFSDDGenerator, TrimSilence
from torchaudio.transforms import MFCC
from torchvision.transforms import Compose

# Create a transformation pipeline to apply to the recordings
transforms = Compose([
    TrimSilence(threshold=1e-6),
    MFCC(sample_rate=8e3, n_mfcc=13)
])

# Fetch the latest version of FSDD and initialize a generator with those files
fsdd = TorchFSDDGenerator(version='master', transforms=transforms)

# Create a Torch dataset for the entire dataset from the generator
full_set = fsdd.full()
# Create two Torch datasets for a train-test split from the generator
train_set, test_set = fsdd.train_test_split(test_size=0.1)
# Create three Torch datasets for a train-validation-test split from the generator
train_set, val_set, test_set = fsdd.train_val_test_split(test_size=0.15, val_size=0.15)

A more complete example can be found here, showing how TorchFSDD can be used to train a neural network.

You can install TorchFSDD using pip.

pip install torchfsdd

Note: TorchFSDD assumes you have the following packages already installed (along with Python v3.6+).

• torch (>= 1.8.0)
• torchaudio (>= 0.8.0)

Since there are many different possible configurations when installing PyTorch (e.g. CPU or GPU, CUDA version), we leave this up to the user instead of specifying particular versions to install alongside TorchFSDD.

Make sure you have torch and torchaudio versions that are compatible!

If you really wish to install torch and torchaudio together with TorchFSDD automatically, the following will install CPU-only versions of both dependencies.

pip install torchfsdd[torch]

Please see the contribution guidelines to see installation instructions for contributing to this repository.

Documentation for the package is available on Read The Docs.

All contributions to this repository are greatly appreciated. Contribution guidelines can be found here.

eonu	black-puppydog

TorchFSDD © 2021-2023, Edwin Onuonga - Released under the MIT License.
Authored and maintained by Edwin Onuonga.