Neural networks for feature computations in automatic speech recognition

Neural networks (NNs) are used for defining good features to use in automatic speech recognition. Of several approaches investigated, the best results were obtained using a NN as a memoryless nonlinear transformation to transform acoustic speech features to a continuous-valued phonetic feature space. The goal is to use labeled training data to automatically derive features which will enhance machine speech recognition. The transformed features were experimentally tested in a syllable recognition task using a hidden Markov model for speech recognition. Syllable recognition rates using the NN-derived features were comparable to those obtained using features derived from a linear transformation