Speech Bandwidth Extension Using Articulatory Features

Abstract

In this thesis, we present a technique for bandwidth extension (BWE) of a narrow-band (0 - 4 kHz) signal using articulatory features. The proposed technique recovers high-band components (4 - 8 kHz) through Gaussian mixture regression (GMR) on both the acoustic and articulatory features from the X-ray Microbeam (XRMB) speech production database. The Gaussian mixture model (GMM) that is based on acoustic and articulatory features is initialized using k-means and iteratively trained using the expectation-maximization (EM) algorithm. BWE experiments were run using data files from different speakers in the XRMB database as train and test data. Time-frequency plots of speech recovered by different training methods are presented in order to show that articulatory trajectories are helpful in characterizing high-frequencied consonants in speech. Finally, we confirm our hypothesis that using GMM with articulation gives better recovery rate is true by performing Student’s t-test on SNR data between original and recovered speech.