Brain-Wave
Classification of Musical Stimuli
Two pilot EEG
experiments to classify brain-wave representations of pitch, timbre, and
loudness.
Blair Bohannan
Music 220C
Spring 2008
Introduction
Two
pilot experiments (one dealing with musical intervals, the other with short
melodies) were designed to investigate the form-bearing dimensions of pitch,
loudness, and timbre, and the role they play in classification rates.
In
these experiments, subjects’ brain waves were recorded while the subjects
listened to auditory stimuli. Classification
using machine-learning algorithms resulted in the rates reported here. Stimuli that classify better (with higher
rates) can be considered more different from each other than those with lower
rates.
Stimuli
Four
stimuli were constructed for each experiment.
They were programmed in ChucK
and recorded using Audacity.
The
interval experiment stimuli vary by four dimensions: interval, timbre,
register, and loudness. The melody
experiment stimuli vary by melody, timbre, register, and loudness. Dimensional
differences are such that the stimuli in the six possible pair combinations
differ within the pair by either two or four dimensions. Click here
for the list of dimensional differences between stimuli pairs.
Click
here for more information on the interval
stimuli.
Click
here for more information on the melody
stimuli.
Hypothesis
Stimuli
pairs that vary by more dimensions (ex. four dimensions of difference versus two)
will yield greater pairwise classification rates.
Methods
Seven
subjects participated, with four subjects run on each experiment (one subject
participated in both). Click here for subject information.
28
monopolar electrodes in the standard 10-20 electrode placement, plus VEOG and
HEOG electrodes, were used for these experiments. All electrodes were referenced to the right
earlobe.
Subjects
were presented a total of 1200 stimuli trials at random (300 of each stimulus). Click here
for more information on experiment setup.
Click
here for filtering information.
Results
Classification
was done two ways: all four stimuli against each other (4-class rates) and all
possible pair combinations.
Classification rates were obtained from a 10-fold cross-validation.
Overview of results – including
4-class and pair rates
Confusion matrices and conditional
probabilities
Contrary
to the hypothesis, stimuli pairs that varied by all four dimensions yielded
both the highest and lowest classification rates. In fact, the lowest pair rate in the
intervals experiment was lower than the 4-class rate. However, classification rates overall were
high and showed a consistent trend in pairwise success rates across all
subjects.
Schedule
Lab
book is here.
Links
http://suppes-brain-lab.stanford.edu
Contact:
blairbo@stanford.edu