Articulatory Inversion of Spanish Speech Signals by means of Machine Learning techniques

The speech inverse problem focuses on retrieving the best articulatory vectors for the task of synthesizing a target voice signal. In this research the problem is approached from a neuromotor level, and by means of some machine learning techniques. Specifically, speech articulators are represented on the midsagittal plane, and controlled by muscular activations grouped in the articulatory vector. The movement of the tongue, by effect of muscular contraction, is derived from a TSK Fuzzy Inference System. On its side, an Echo State Network is used to model the glottal airflow. Later, Continuous Genetic Algorithms evolve populations of articulatory configurations in order to approximate acoustic features of target spanish vowels, and of consonants /m/, /n/, /f/ and /s/. The recovered midsagittal configurations along with subjective tests performed by a group of evaluators, positively verify the effectiveness of these techniques. Key–Words: Speech inverse problem, neuromotor inversion, midsagittal model, fuzzy logic, echo state networks, genetic algorithms.