What it does
MagicVoice is an app to add voice style to your speech. Like Prisma for photos, MagicVoice uses neural style transfer to learn the sound characteristics of pop stars, and transform the audio with the style.
How we built it
We built the frontend with iOS, and backend with tensorflow.
Challenges we ran into
To apply neural style transfer, we need a pre-trained model that's good enough. For vision, there are plenty of existing pre-trained convolutional neural network models. But we found it challenging to use an open-source audio model.
What's next for MagicVoice
We will train the audio model from scratch using convolutional neural networks on a large corpus. Once the model is obtained, we can proceed with the work on style transfer.