In recent years, many methods have been introduced for supporting the diagnosis of stuttering for automatic detection of prolongation in the speech of people who stutter. However, less attention has been paid to treatment processes in which clients learn to speak more slowly. The aim of this study was to develop a method to help speech-language pathologists (SLPs) during diagnosis and treatment sessions. To this end, speech signals were initially parameterized to perceptual linear predictive (PLP) features. To detect the prolonged segments, the similarities between successive frames of speech signals were calculated based on correlation similarity measures. The segments were labeled as prolongation when the duration of highly similar successive frames exceeded a threshold specified by the speaking rate. The proposed method was evaluated by UCLASS and self-recorded Persian speech databases. The results were also compared with three high-performance studies in automatic prolongation detection. The best accuracies of prolongation detection were 99 and 97.1% for UCLASS and Persian databases, respectively. The proposed method also indicated promising robustness against artificial variation of speaking rate from 70 to 130% of normal speaking rate.
In this paper, we continue our previous work on nonlinear feature compensation of distortions in clean and telephone speech recognition systems. We have shown that Bidirectional Neural Network (Bidi-NN) can compensate nonlinearly-distorted components of feature vectors. In this study, we present a new effort to improve recognition accuracy on clean and telephone speech data by employing a two-stage feature compensation technique for recovering optimal (from a classification point of view) Log-Filter Bank Energies (LFBE). These new features are achieved by training a new Bidi-NN with compensated features and considering compensated feature as the input data to Bidi-NN. We also achieved MFCC features by applying discrete cosine transform (DCT) to compensated Log-Filter Bank Energies (LFBE) features. HMM phone models are trained on these modified features. By using the two-stage compensated features, we obtained an absolute improvement of 4.73% and 9.29% in phone recognition accuracy compared to baseline system in clean and telephone conditions respectively. We also obtained an absolute improvement of 25.67% in phone recognition accuracy for the system which was trained on clean data but tested on telephone data. These results show excellency of NN-based nonlinear compensation of speech feature vectors in HMM-based speech recognition systems.
We showed in our recent work that Bidirectional neural network (BNN) is a powerful tool for feature compensation in automatic speech recognition systems. In this paper, we have introduced BNN as feature compensator for better discriminating of pathological voices from normal subjects. Mel-Frequency Cepstral Coefficients (MFCCs) were extracted from each frame of sample voices and were compensated in two steps. First, BNN is trained with both normal and pathological feature vectors. Our hypothesis is that BNN can extract useful knowledge about the patterns of each class during training step. In second step, MFCC feature vectors feed into BNN and compensate according to latent knowledge of BNN. In the last step, Compensated MFCCs are classified as pathological or normal by HMMs. We achieved 4.67%, 2.81% and 2.24% improvement in measures of specificity, accuracy and sensitivity by compensated feature vectors compared to the original feature vectors. Results corroborated our hypothesis about the ability of BNN in compensation of feature vectors in a way that these features become more suitable for detection of pathological voices from normal ones.