ImageNet Voice

This dataset is derived from ImageNet, and contains synthetic recordings of TTS voices saying each ImageNet label 200 times. This is more or less a toy dataset, but can be very useful for projects with little-to-no data.

Usage

After installing with python3 -m pip install imagenet-voice, ImageNet-Voice can be imported and used like most other datasets:

import imagenet_voice

X, y = imagenet_voice.load(progress=True, download_data=True)

The Python package includes 3 loading functions:

• load(progress, download_data) - Load the entire dataset. If progress, display a progressbar. If download_data, git clone the dataset.
• load_classes(classes, progress, download_data) - Loads only the classes in the provided classes list.
• load_n(n, random, progress, download_data) - Loads n classes, alphebtically unless random=True.

Content

This dataset is currently at 28% capacity, and more will be uploaded soon. Each folder, representing a category in ImageNet, contains 200 unique TTS files generated using ttsddg using the 7 pre-installed voices in OSX.

Acknowledgements

pyttsx3 was integral to creating ttsdg. As much trouble as I had with ffmpeg and other audio libraries, pyttsx3 made it much easier than any other option I could've taken to generate this dataset.

I appreciate any help that anyone would be willing to offer to get this dataset at 100% capacity any sooner. Please contact me!

Inspiration

I created this dataset to provide training data for a GAN that would take in an audio file of a spoken word and would generate an image of whatever was spoken. I still plan to work on that project, but I'd like to see other people perform similar tasks.