Acoustic Space Movement Planning in a Neural Model of Motor Equivalent Vowel Production

Abstract

Recent evidence suggests that speakers utilize an acoustic-like reference frame for the planning of speech movements. DIVA, a computational model of speech acquisition and motor equivalent speech production, has previously been shown to provide explanations for a wide range of speech production data using a constriction-based reference frame for movement planning. This paper extends the previous work by investigating an acoustic-like planning frame in the DIVA modeling framework. During a babbling phase, the model self-organizes targets in the planning space for each of ten vowels and learns a mapping from desired movement directions in this planning space into appropriate articulator velocities. Simulation results verify that after babbling the model is capable of producing easily recognizable vowel sounds using an acoustic planning space consisting of the formants F1 and F2. The model successfully reaches all vowel targets from any initial vocal tract configuration, even in the presence of constraints such as a blocked jaw.