Spatial and channel attention mechanism for speech disfluency detection using deep learning technique

Abstract

Stuttering is a speech communication disorder, it is characterized by repetitions, prolongation, and unusual pauses that cause interference with the natural flow of speech. In recent times, automatic speech recognition and speech processing systems have gained enormous attention because they are used in most of the human machine interaction applications. However, the performance of these systems is affected by stutter speech, stutter detection is the major challenge due to speech disfluencies. To address this major challenge, this paper introduced a novel deep learning (DL) based paradigm, which integrates a hybrid feature extraction algorithm, with the Spatial and Channel attention mechanism to refine the features and for reliable detection of speech disfluency. This study is conducted on multiple stutter data set which includes UCLASS (Release 1, Release 2), FluencyBank and SEP-28k. The major drawback of all these data sets is data imbalance. To reduce this imbalance, the author used data augmentation techniques, which includes, noise, music, reverberation and pitch shifting methods. However, increasing the stutter detection accuracy remains a challenging issue. To address this issue, the author proposed a hybrid feature extraction model, which extracts temporal, contextual, spectral, and pitch information from the speech signal. The obtained features are then processed through the attention mechanism where channel and spatial attention models help to refine the features. Finally, a multiclass convolutional neural network (CNN) classifier is used to detect the stutter event in the speech signals. The results show that our model with spatial and channel attention mechanism performs better than existing deep learning approaches and accurately detects stuttering.