Improved automatic speech recognition system using sparse decomposition by basis pursuit with deep rectifier neural networks and compressed sensing recomposition of speech signals

Research on the common limitations of Automatic Speech Recognition (ASR) systems state problems ranging from environmental noise, and channel or speaker variability to the limitations imposed by the measurement device. In mobile applications for automatic speech recognition, the Nyquist criteria imposes more limitations on the sampling rate at which a device can acquire signal, often the lack of fidelity of the acquired signal causing bad speech recognition. This is a specific problem for mobile devices (which are also, nowadays, the prime beneficiaries of speech recognition applications) as in this case the sampling rate is limited. We envisage a way to get the best out of any acquired signal, by use of sparsity decomposition algorithms and compressed sensing recomposition. We focus on the fact that complex sounds can be viewed as an overlapping of a number of sounds coming from simple sparse sources. Therefore, we decompose the measured signal in a linear combination of simple sparse signals and we reconstruct each sparse signal by means of compressed sensing recomposition in order to gain a better signal fidelity. We make use of deep rectifier neural network designed to decompose a training set of signals and compute a specific dictionary with simple sparse signals. The resulted sparse signals are used for decomposing the acquired signal by means of sparse algorithms, and, consequently, the resulted combination of sparse signals will be used for signal reconstruction in a compressed sensing algorithm. We test the framework for different simulated speech signals, as well as its usability in automatic speech recognition, discussing the improvements this upgrade brings to an ASR. In this paper we will describe the framework and the algorithms used and present the experimental results.