Latent Birds: A Bird's-Eye View Exploration of the Latent Space

The use of a generative approach for sound synthesis breaks
through the limitations of traditional approaches, proposing novel ways to explore creative ideas. This paper demonstrates a method to generate original bird vocalizations using a Variational Convolutional Autoencoder trained on
mel-spectrograms of bird song and call recordings. The
vocalizations are reconstructed by sampling the latent space
and decompressing the resulting mel-spectrogram. The
results are quite promising, in that our system is able to
generate a variety of bird vocalizations depicting plausible
songs and calls, by interpolating between existing vocalizations or sampling the latent space. A Twitter bot that
publishes a unique daily bird vocalization is also implemented.