Silent Speech Interface: An Inversion Problem

Abstract

When conventional acoustic-verbal communication is neither possible or desirable, silent speech interfaces (SSI) rely on biosignals, non-acoustic signals created by the human body during speech production, to facilitate communication. Despite considerable advances in sensing techniques that can be employed to capture these biosignals, majority of them are used under controlled scenarios in laboratories. One such example is Electromagnetic Articulograph (EMA), which monitors articulatory motion. It is expensive with inconvenient wiring and practically not portable in real world. Since articulator measurement is difficult, articulatory parameters may be estimated from acoustics through inversion. Acoustic-to-articulatory inversion (AAI) is a technique for determining articulatory parameters using acoustic input. Automatic voice recognition, text-to- speech synthesis, and speech accent conversion can all benefit from this. However, for speakers with no articulatory data, inversion is required in many practical applications. Articulatory reconstruction is more useful when the inversion is speaker independent. Initially, we analysed positional data to better understand the relationship between sensor data and uttered speech. Following the analysis, we built a speaker independent articulatory reconstruction system that uses a Bi- LSTM model. Additionally, we evaluated the trained model using standard evaluation measures.