Making our Voices Heard: The Art of Machine Learning for Speech Classification

Emerging voice activated technologies depend on fast and accurate classification of a coustic s ignals. R ecent advances in machine learning and application specific computational systems have enabled highly accurate classifiers when trained with a large amount of data. This research presents the performance of a convolutional neural network when trained with distinct feature sets and assesses the performance tradeoff in size and feature-selective training. The experiments are conducted using the Free Spoken Digit Dataset that includes spoken digits from six different speakers with different accents. The classification e fficiency of on e an d tw o la yer CN N’s are reported using three different types of spectrograms. To encourage the participation of a student body in machine learning applications of acoustics, the essential features of the neural network are captured in digital art form to render a more accessible introduction to the topic.