Perceptually Controlled Reshaping of Sound Histograms

Many audio processing algorithms have optimal performance for specific signal statistical distributions that may not be fulfilled for all signals. When the original signal is available, we propose to add an inaudible noise so that the distribution of the signal-plus-noise mixture is as close as possible to a given target distribution. The proposed generic algorithm (independent from the application) adds iteratively a low-power white noise to a flat-spectrum version of the signal, until the target distribution or the noise audibility is reached. The latter is assessed through a frequency masking model. Two implementations of this sound reshaping are described, according to the level of the targeted transformation and to the foreseen application: histogram global reshaping (HGR) to change the global shape of the histogram and histogram local reshaping (HLR) to locally “chisel” the histogram, but keeping the global shape unchanged. These two variants are illustrated by two applications, where the inaudibility of the noise generated by the algorithm is required: “sparsification” for source separation and low-pass filtering of the histogram for application of the quantization theorem, respectively. In both cases, the target histogram is reached or almost reached, and the transformation is inaudible. The experiments show that the source separation performs better with HGR and that the HLR allows a better application of the quantization theorem.