Predicting Perceptual Centers Located at Vowel Onset in German Speech Using Long Short-Term Memory Networks

Perceptual centers (p-centers) can be defined as the perceived centers of a syllable. Previous research regarding the location of p-centers in speech relied on experimental methods, and among the suggested acoustic features contributing to the location of p-centers in Germanic languages is the transition of the consonant to the vowel onset. The current study investigates the prediction of the location of p-centers in German, by means of machine learning. Machine learning is a promising tool to capture possible non-linear relationships that may occur among the acoustic features used in the complexity that is the human perception. Therefore, an LSTM neural network approach was used for the identification of p-centers in a set of spoken German sentences, with input data features being Mel Frequency Cepstral Coefficients (MFCC), amplitude envelope and root mean squared energy. The model was able to achieve a balanced accuracy of 84% with MFCCs being the best predictor of p-center location.