Introduction
Nowadays, music information retrieval (MIR) is a very popular and growing
field with many real-world applications. Many people use apps for automatic
songs recognition daily. But even more untrivial tasks such as genre
recognition, personalized recommender systems or music generations are
part of the research in MIR. The biggest challenges are high dimensional
representation and feature extraction. In our project, we summarize some of
the most interesting applications of MIR and examine several state-of-theart
machine learning approaches.
Emotion Recognition
Composer Classification
D A T A M I N I N G I N M U S I C
Music source separation
Musical instrument identification is a crucial problem in music information
retrieval. Without it, we would not be able to: search music by instrument,
recognize musical genres, or make accurate music transcription. Music
source separation algorithms have also been successfully used in
predominant pitch tracking, accompaniment generation for Karaoke or
singer identification. Modern techniques involve methods such as “deep
clustering” or “deep convolutional neural networks”.
Queries for production music (stock music) contain quite a lot of emotional
terms. This creates a need for the music industry to categorize large-scale
online music collections. This need is naturally helping music emotion
recognition field to expand.
At the same time explanations from the prediction model could be used as
parameters in the fields of recommender systems or music generation. The
most common methods used are Convolutional Neural Networds, Long ShortTerm
Memory Recurrent Neural Network.
Automatic music transcription
Automated creation of human readable representation of audio wave inputs
is a challenging problem which has implications in many parts of the music
industry, from visualization of music content through music search and
recommendation to helping improve music creation. Current approaches are
based on the usage of both long short-term memory networks and nonnegative
matrix factorization.
Machine learning models which recognize a composer based on the musical
scores or the audio files are extensively studied in the music information
retrieval community. The major challenges in this field are learning from
limited datasets and sufficient data representation. Many approaches have
been taken both in the classification of musical scores - by Convolutional
Neural Networks, as well as in recognizing the composer by audio files - by
Decision Trees, SVM, logistic regression and Neural Networks.
Petr Kadlec · Ondrej Kužlík · Klára Petrovičová · Tomáš Repák · Samuel Šuľan
Music generation
Automatic music generation is a complicated task related to MIR. Nevertheless,
there are already existing algorithms, which are able to produce plausible
music. Since music is a temporal domain, models like Recurrent Neural
Networks, LSTMs, or Gated Recurrent Units are usually used. While these
models are widely employed, the problem of music generation has not been
solved yet.