StutterNet: Stuttering Detection Using Time Delay Neural Network

Robust Stuttering Detection via Multi-task and Adversarial Learning

“…This study is an extension of our previous work Stutter-Net [16]. Due to the diversity and uniqueness of stuttering, it continues to be the most demanding and challenging to detect, due to its inherent nature of huge class imbalance across different types of disfluencies [2].…”

“…The networks were trained using the spectrograms, which were the sole input features to the model. The proposed architecture showed promising results on a small subset of speakers (25) of UCLASS dataset, but did not perform well on a large set of speakers [16]. In addition, they have mentioned it end-to-end, but they are still using the hand crafted spectrogram features.…”

“…Even though the model shows promising results but they have mixed all the datasets together for their case study. We recently introduced a single branch multi-class StutterNet, a time-delay neural network based stuttering classifier, capable of identifying fluent segment and core behaviours [16], and, it shows encouraging results compared to state-of-the-art on a large set of speakers. The StutterNet was trained with MFCCs and cross entropy on a set of 104 speakers from UCLASS dataset.…”

“…Our SD models are based on time delay neural networks, which has been proven effective in various speech domains [16]. The model is fed with mean-normalized 20dimensional MFCC input features extracted on a 20 ms sliding window with a hop length of 10 ms.…”

“…Evaluation Metrics: To test and evaluate the model performance, we use the widely and standard metrics that are used in stuttering and other speech tasks [2] which include recall, precision, F1-score and accuracy. The models are evaluated and compared to our previous work StutterNet [16]. The results reported are the average of 10 experiments with 10fold cross validation technique.…”