Modification of Devoicing Error in Cleft Lip and Palate Speech

Abstract

The cleft of the lip and palate (CLP) caused by structural and functional deformation leads to various speech-related disorders, which substantially degrades the speech intelligibility. In this work, devoiced stop consonants in CLP speech are analyzed, and an approach is proposed for its modification in order to enhance the speech intelligibility. The devoicing errors are primarily characterized by the absence of voicebar in the closure interval and relatively longer voice onset time (VOT). The proposed approach first segments the regions corresponding to the closure interval and VOT based on the knowledge of glottal activity, voice onset point, voice offset point, and burst onset point. In the next stage, specific transformations are performed for the modification of closure bar and VOT respectively. For transformation, first different transformation matrices are learned for closure bar and VOT from normal and CLP speakers. The transformation matrix is optimized using nonnegative matrix factorization method. Further, the corresponding transformation matrices are used to modify the closure bar and VOT separately. The subjective evaluation results indicate that the devoiced stop consonants tend to exhibit the characteristics of voiced stop consonants.

Enhancing the Intelligibility of Cleft Lip and Palate Speech using Cycle-consistent Adversarial Networks

“…We now focus on the results of the studies conducted. The NMF based enhancement method is well explored in previous pathological speech studies [23,34,43]. Here, the speech signals are processed in a frame size of 20 ms and a shift of 10 ms.…”

“…The prior works show the use of GMM based VC and nonnegative matrix factorization (NMF) based VC for the improvement of various pathological speech [16,23,34]. In this kind of method, temporally aligned parallel source (pathological speech) and target (normal/non-pathological speech) are required for training.…”