Signal Processing - July 2017 - 143

SP COMPETITIONS
Craig T. Jin, Matthew E.P. Davies, and Patrizio Campisi

Embedded Systems Feel the Beat in New Orleans
Highlights from the IEEE Signal Processing Cup 2017 Student Competition

F

oot-tapping and moving to music is
such a natural human activity, one
may assume that feeling the beat in
music is a simple task. Feeling the beat
and then producing it, e.g., by foot tapping, is an intrinsically real-time process.
As listeners, we do not wait for the beat
to occur before tapping our foot; instead,
we make predictions about when the next
beat in the music will occur and continually revise our sense of the beat based on
the accuracy of our predictions. Likewise, performing musicians have shared
sense of beat, which is what allows them
to play in time together.
This type of high-level music listening and understanding sits at the heart of
the challenge set for this year's IEEE
Signal Processing Cup (SP Cup) competition, the final stage of which concluded
at the 2017 IEEE International Conference on Acoustics, Speech, and Signal
Processing (ICASSP), hosted in New
Orleans, America's jazz heartland, on
5 March. The participating undergraduate
cohort had to devise and construct a creative, embedded application demonstrating a real-time response to the beat of the
music. Depending on the genre, composition, and rhythmic complexity of a
musical piece, real-time beat tracking
poses considerable challenges, which are
equally present for human listeners, especially those without formal musical training. Throughout the SP Cup, the teams
Digital Object Identifier 10.1109/MSP.2017.2698075
Date of publication: 11 July 2017

1053-5888/17©2017IEEE

confronted these challenges from both
the human and computational perspectives via the choice of training and testing
material, the human annotation of beat
locations, the implementation and evaluation of their beat-tracking algorithms,
and the response to the beat in their creative applications.

Beat tracking in music signals
The task of beat tracking of music signals has been an active area of music
signal processing research for more than
25 years. While many of the earliest
computational approaches sought to
emulate the human process of tapping
the beat in real time by making predictions of future beats [1], [2], a marked
shift occurred in the early-to-mid 2000s
toward offline approaches that could
observe the entire musical input prior to
determining beat locations.
The standard pipeline for offline beat
tracking involves the explicit identification of note onset locations (or an "onset
strength function," which emphasizes
their location) that are subsequently
passed to a tempo-estimation stage used
to estimate the latent beat periodicity in
the input signal, followed by the recovery of the phase (or alignment) of the
beats to the music. Common techniques
used to extract the beat from music signals include multiagent systems, dynamic programming, hidden Markov
models, and a mixture of experts sys tems. Current state-of-the-art methods
employ deep neural network architecIEEE SIGNAL PROCESSING MAGAZINE

|

July 2017

|

tures to learn the relationship between
labeled beat annotations in training data
sets and feature representations extracted
from musical audio signals, thus leveraging both advanced signal processing and
machine learning.
The growth of offline approaches
arose in part by the significant increase
in the use of beat tracking for so-called
beat-synchronous analysis as an intermediate processing step within other music
signal analysis tasks, such as structural
segmentation, chord detection, and
music transcription. With the shift
toward making multiple passes across
input signals, the focus on real-time
analysis was reduced. Furthermore, with
a greater emphasis on the accuracy of
beat tracking over computational efficiency, offline approaches also provided the opportunity for tracking the beat
in music with expressive timing (i.e.,
changes in tempo) something that was
considered impossible for real-time systems bound by the need to make predictions of future beats in the music [3].
An emerging topic related to the
domain of music signal processing is creative music information retrieval, which
seeks to open new possibilities for music
creation, interaction and manipulation.
This is facilitated by the robust analysis
and interpretation of music signals [4].
For applications that target live interaction between users and/or musicians and
technology, there is a compelling need to
perform music signal analysis in real
time. One specific motivation for the SP
143
Table of Contents for the Digital Edition of Signal Processing - July 2017
