A Comparison of Speech-to-Speech Neural Network Methodologies for Digit Pronunciation

In this work, the classical problem of digit recognition and pronunciation from an audio source in Spanish is revisited and compared with that of directly teaching a deep neural network to pronounce the corresponding digit. While the first approach roughly corresponds to that of most current speech processing methodologies that intend to identify and reconstruct phonetic units, before performing any task from reproducing to translation, the second approach is rarely found in the literature despite the fact that it is clearly more biologically inspired than the first one. Advantages and disadvantages of both methodologies are discussed based on the obtained results.