Modern synthesizers are getting increasingly powerful and now provide an overwhelming amount of parameters to carve a sound spectrum. This simultaneously increases creative freedom but can also complicate the sound design process. In parallel, recent generative learning models have been developed towards audio synthesis. Here, we aim at providing intuitive control over sound synthesis with deep learning models, through synthesis by learning. Only a limited number of approaches have been proposed to deal with this new type of synthesis, which allows learning a synthesizer directly from audio sample examples. One of the most important proposals relies on the framework of variational autoencoders, which allows generating sounds from a parameter latent space, by simultaneously learning inference and generation networks from existing data. In this work, we develop generative models for audio synthesis that are able to handle complex temporal information, thus, allowing to generate a wide variety of sounds. Here, we developed a model based on combinations of variational autoencoders and convolutional neural networks for audio synthesis. We also collected and labeled a dataset representing a variety of percussive sounds to train our model.