Outcomes of Deep Neural Network Hyperparameter Tuning on Bengali Speech Token Classification

Speech-based human-computer interaction is a vital research area where Bengali speech requires substantial research for being efficient like English. Nowadays, most speech-related research involves Deep Neural Networks (DNNs). We present comprehensive results on DNN hyperparameter tuning on isolated Bengali vowel and word classification. Two separate datasets of isolated Bengali vowels and words are utilized in a four-hidden-layered DNN configuration. Five formant frequencies have been extracted for all fourteen vowels and words, and after applying a necessary feature standardization technique, these acoustic features are utilized in the DNN input layer. Three hyperparameters–optimizer, batch size, and dropout–have been tuned to develop an optimum DNN classification model. We report the outcome achieved after applying different sets of values for these three hyperparameters. Finally, we recommend an Adadelta optimizer, a batch size of 128 or 256, and a dropout between 0.25 to 0.4 after the first hidden layer. The DNN formed by utilizing these recommended hyperparameters will be a potential pre-configured speech classifier for various humancomputer interaction-based devices and systems.