Deep Optimization of Parametric IIR Filters for Audio Equalization

This paper describes a novel Deep Learning method for the design of IIR parametric filters automatic multipoint audio equalization, that is task improving sound quality listening environment at multiple points employing loudspeakers. The are designed to approximate inverse RIR and achieve almost flat magnitude response. A simple effective neural architecture, named BiasNet, proposed determine equalizer parameters. architecture conceived optimization and, as such, able produce optimal parameters its output, after training, with no input required. In absence input, presence learnable non-zero bias terms ensures network works properly. An output scaling used obtain accurate tuning center frequency, factor gain. All layers involved in shown be differentiable, allowing backpropagation optimize weights achieve, number training iterations, according given RIR. optimized respect loss function based on spectral distance between measured desired response, regularization term keep same microphone-loudspeaker energy balance equalization. Two experimental scenarios employed, room car cabin, several performance improves over baseline techniques achieves an band lower computational cost.