Audio Bigrams as a Unifying Model of Pitch-based Song Description

Abstract

In this paper we provide a novel perspective on a family of music description algorithms that perform what could be referred to as `soft' audio fingerprinting. These algorithms convert fragments of musical audio to one or more fixed-size vectors that can be used in distance computation and indexing, not just for traditional audio fingerprinting applications, but also for retrieval of cover songs from a large collection, and corpus-level description of music. We begin with a high-level overview of the algorithms. Next, we identify and formalize an underlying paradigm that allows us to see them as variations of the same model. Finally, we present pytch, a Python implementation of the model that accommodates several of the reviewed algorithms and allows for a variety of applications. The implementation is available online and open to extensions and contributions.