Gamma Boltzmann Machine for Audio Modeling

This paper presents an energy-based probabilistic model that handles nonnegative data in consideration of both linear and logarithmic scales. In audio applications, magnitude time-frequency representation, including spectrogram, is regarded as one the most important features. Such magnitude-based features have been extensively utilized learning-based processing. Since a scale terms auditory perception, are usually computed with function. That is, function applied within computation so learning machine does not to explicitly scale. We think different way propose restricted Boltzmann (RBM) simultaneously models linear- log-magnitude spectra. RBM stochastic neural network can discover representations without supervision. To manage scales, we define energy based on results conditional distribution (of observable data, given hidden units) written gamma distribution, hence proposed termed gamma-Bernoulli RBM. The was compared ordinary Gaussian-Bernoulli by speech representation experiments. Both objective subjective evaluations illustrated advantage model.