Speech disfluency recognition using convolutional neural networks (CNNs)

Abstract

Abstract- This paper presents a comparative study of applying different convolutional neural network (CNN) architectures for speech disfluency recognition. Furthermore, the paper presents a comparative study of three different types of features extracted from the speech clip under examination. These features are the spectrogram computed by the Short-Time Fourier Transform (STFT), the Mel spectrogram and the Mel-Frequency Cepstral Coefficients (MFCCs). Four CNN architectures, ResNet50, VGG16, LeNet-5, and AlexNet, were evaluated on a speech database developed by the LANNA research group, consisting of recordings from normal children (normal speech) and children with specific language impairment (SLI) (stuttered speech). The performance of each network for each of the three feature types was assessed using the classification accuracy and the area under the receiver operating characteristic curve (ROC-AUC). Results showed that AlexNet and LetNet5 provide the best performance with the STFT-based spectrogram. Both Resnet50 and VGG16 provide the best performance with the MFCCs feature.