Digital Audio Signal Processing, 3rd Edition

Chapter 11Machine Learning for Audio

P. Bhattacharya P. Nowak and U. Zölzer

11.1 Introduction

Machine learning can be described as a set of methods or algorithms that offer the capability to learn from data automatically and develop flexible parameterized models. In classical machine learning, the important features of the input are manually designed and extracted from data, and the system automatically learns to map these features to the requisite outputs. This learning is usually achieved through parameter adaptation and/or hyper‐parameter tuning based on a predefined goal or optimization of a cost function. Classical machine learning is used in, and works well for, simple pattern recognition problems. A bulk of the effort or computation would include the design of optimal features for the system. Once the features are handcrafted from a dataset, a generic classification or regression model is used to obtain the output. Some examples of classical machine learning algorithms include linear regression, logistic regression, k‐nearest neighbors, and simple decision trees. Representation learning goes a step further and mostly eliminates the need for handcrafted features. Instead, the required features are automatically discovered from the data. With the evolution of graphical processing units, new machine learning methods supporting more complex models have come to the forefront, with deep learning being one of the major topics of research. Deep learning is part of a broader ...