Prosody Tutorial: Lecture 19: Features for Machine Learning

This is Video 19 of our series on prosody. Systems for many tasks --- such as speech recognition, emotion recognition, and intent recognition --- are built today by applying machine learning methods to large datasets. The input to these models is usually some set of features computed from the audio signal. In this lecture we discuss the advantages and disadvantages of various types of prosodic features for various purposes. 00:00 Machine Learning needs Features 00:49 Different Types of Features 01:13 Using Meaningful Features 02:00 Using Midlevel Features 02:58 Using Frame-Level (Low-Level) Features 03:55 Using Filterbank and other Generic Features 04:47 Using Features from Pretrained Models 06:41 Feature Set Choices for Common Tasks 07:10 Summary