The Autocoder package is a tool based around a variational autoencoder––a neural network capable of learning a spectral representation of a soundfile and synthesizing a novel output based on the trained model.

A spectral representation is extracted from an input sound by running it through an STFT analysis, producing a number of spectral frames representing the sound moment by moment. The software through an encoder that compresses the data into a latent layer. The latent layer has eight values, each encoding some aspect of the input data. By feeding the training data through the encoder after training. By feeding arbitrary values as input into the latent vector layer, the encoder returns a new spectral frame that represents an unseen point within the spectral space of the training data and can then be used in any number of ways, e.g. for synthesis or convolution, as an impulse response in a hybrid reverb, or for cross–synthesis.

The package provides a simple and easily extendable ecosystem to assist with the experimentation and development of sound software and hardware based on the underlying neural network architecture. It is available both in code and in hardware form and comes with an osc interface allowing for easy integration with Max/MSP.