DANN: Digital Audio Neural Network

Abstract

Neural networks have been widely used in computer audio to deal with many synthesis and processing parameters. Experiments have also led to direct synthesis of audio using dynamic neural networks that exhibit oscillating behaviour, as they are clocked at audio rates. There is clearly great potential in convolutional networks as they directly implement complex resonances enabling acoustic interaction. It thus transpires that many different neural network types were used for direct audio synthesis and processing. The proposed innovation extends a fully-connected network of computational cells to facilitate long feedback delays (needed for waveguides) intertwined with zero delays (needed for polynomials). Another improvement is the network computing algorithm that allows for randomisation of delays by adjusting the order of calculations. The DANN model is a specification for network of cells and links that allows direct translation from digital audio synthesis and processing schemes. Taking basic examples such as exponential decay, logistic chaos and biquad filter it transpires that there are exact DANN equivalents to these basic building blocks. It is therefore suggested that complex DANN schemes can reproduce myriad synthesis and processing techniques in a transparent fashion. As a matter of proving the concept a simple snare drum synthesizer is translated to a DANN scheme in order to explore the parameter space in this domain. Thanks to the large amount of training schemes there is great potential be found in training networks for velocity sensitive synthesis to include complex physical interaction sensors. Audio signal processing applications may include modelling of analogue equipment and symbolic acoustic instrument processing. The simplicity of the network grants easy implementation on fast architectures to facilitate near real-time processing and higher sample rates. Beyond the practical uses proposed it is further suggested that network simplification procedures may lead to yet unknown compact schemes for complex signal processing problems in the audio domain, potentially reverberation and pitch-shifting.