The overwhelming size of musical data online compels the need for tools to efficiently access and organize these songs. However, there is an existing gap between the highly mathematical, impersonal recommendations and the recommendations embodying the actual perceptions of a user. We want to create an efficient recommendation program that accurately represents a user’s musical preferences.
Current music recommendation systems include Pandora and
Last.fm. Pandora uses a team of music analysts to tag songs, which leads to the
possibility of introducing human error from the analysts. Last.fm uses
user-generated tags to determine song similarity; these tags are often
ambiguous and do not take into account the quantifiable content from a song’s
musical data.
It is clear that people do not perceive music by the
representation of numerical similarity; therefore, we must consider abstract
features that model the perception of music.
We have crafted our research questions based on the areas that we
believe require growth to advance the field of music recommendation software:
How can music perception
be incorporated into an automated recommendation system that focuses on an individual’s music preferences
and listening trends?
Our question takes into account how listeners perceive music and
combines that with quantifiable, automated characteristics from audio files.
In response to this question, our hypothesis is:
If we can computationally
analyze qualities and trends within
a specific listener’s song list, then we can generate a model of the user’s perception of music that will account for abstract factors like mood.
Originally, our methodology focused on the development of Matlab
code to extract features from a song as described in our Thesis Proposal. Features are mathematically defined
components within the music, such as tempo or key signature. In feature
extraction of audio signals, you start with a song file, find the corresponding
audio signal and then extrapolate certain data points, or features, from this
waveform. We were going to use music theory to determine which
higher-order features we would extract from this type of data, and then compare
the features of songs to each other to estimate the degree of musical
similarity for song recommendation.
However, we have
recently realized that this method, being so mathematically oriented, seems
rather impersonal from a user’s perspective as it fails to take into account
the individuality of the user. We have been working on refocusing
our approach, so that we could allow our program to be more attentive to the
user’s individual musical preferences.
This change in approach led our search for APIs that would
generate musical features, which we can then apply to music recommendation in a
way that models a user’s musical cognition. The use of an API allows us to
concentrate on the user-specificity of the program, especially in dealing with
the perception of music. We found The Echo Nest, which is an API that returns
numbers corresponding to an input song’s features. We will develop a cognitive
music model in order to represent the musical perception of listeners and then
choose features from The Echo Nest’s API for our program.
The overall outline of our project is as follows:
(1) Compilation of research on current music recommendation
systems, feature generation, The Echo Nest’s API, and statistical analysis
methods. Although this is listed only as a first step, this is an ongoing
action.
(2) Experimentation with feature generation and contact of
experts based on our testing and research proposal.
(3) Use of Focus Group sessions to determine satisfaction levels
of current music recommendation systems and elements of songs on which people
focus.
(4) Testing of the The Echo Nest API for use as a feature
generation engine.
(5a) Creation and release of an online survey to compare human
perception features of songs (from the Five-Factor Model Experiment) to
features from The Echo Nest API.
(5b) Comparison of the Five-Factor Model and The Echo Nest’s API
features to find correlations using statistical methods such as the ANOVA test.
(6) Prediction of the Five-Factor Model from the API
correlations made in (5b).
(7) Generation and release
of final online survey to observe if similarities between the Five-Factor Model
and The Echo Nest’s API led to valid recommendations.
(8) Analysis of survey in (7).
