“A Learning Problem”: Deep Learning Target Word Classification for Infant Directed Speech

Abstract

Target word classification is an important part of speech recognition and artificial intelligence. Due to the rise of deep learning, target word classification models have shown impressive results. However, for some types of data, the accuracy of these models is still lacking. For one, research has found that using state-of-the-art target word classification software for Infant Directed Speech (IDS) - a special type of speech used when talking to infants - results in lower classification accuracy compared to Adult Directed Speech (ADS). In this thesis, we will answer the question: "Can deep learning models be used for successful classification of target words in Infant Directed Speech?" To answer this question two experiments have been conducted in which deep learning classification models (CNNs and RNNs) were trained and evaluated on IDS and ADS. The results of these models have been compared and analyzed. There was found to be no significant difference in classification accuracy between the two types of speaking. Furthermore, the CNN model classified IDS test samples with an accuracy of 85%. From this, it was concluded that deep learning models can be used for successful target word classification of IDS.