Music information retrieval


Music information retrieval is the interdisciplinary science of retrieving information from music. MIR is a small but growing field of research with many real-world applications. Those involved in MIR may have a background in musicology, psychoacoustics, psychology, academic music study, signal processing, informatics, machine learning, optical music recognition, computational intelligence or some combination of these.

Applications

MIR is being used by businesses and academics to categorize, manipulate and even create music.

Recommender systems

Several recommender systems for music already exist, but surprisingly few are based upon MIR techniques, instead making use of similarity between users or laborious data compilation. Pandora, for example, uses experts to tag the music with particular qualities such as "female singer" or "strong bassline". Many other systems find users whose listening history is similar and suggests unheard music to the users from their respective collections. MIR techniques for similarity in music are now beginning to form part of such systems.

Track separation and instrument recognition

Track separation is about extracting the original tracks as recorded, which could have more than one instrument played per track. Instrument recognition is about identifying the instruments involved and/or separating the music into one track per instrument. Various programs have been developed that can separate music into its component tracks without access to the master copy. In this way e.g. karaoke tracks can be created from normal music tracks, though the process is not yet perfect owing to vocals occupying some of the same frequency space as the other instruments.

Automatic music transcription

Automatic music transcription is the process of converting an audio recording into symbolic notation, such as a score or a MIDI file. This process involves several audio analysis tasks, which may include multi-pitch detection, onset detection, duration estimation, instrument identification, and the extraction of harmonic, rhythmic or melodic information. This task becomes more difficult with greater numbers of instruments and a greater polyphony level.

Automatic categorization

Musical genre categorization is a common task for MIR and is the usual task for the yearly Music Information Retrieval Evaluation eXchange. Machine learning techniques such as support vector machines tend to perform well, despite the somewhat subjective nature of the classification. Other potential classifications include identifying the artist, the place of origin or the mood of the piece. Where the output is expected to be a number rather than a class, regression analysis is required.

Music generation

The automatic generation of music is a goal held by many MIR researchers. Attempts have been made with limited success in terms of human appreciation of the results.

Methods used

Data source

give a clear and logical description of music from which to work, but access to sheet music, whether digital or otherwise, is often impractical. MIDI music has also been used for similar reasons, but some data is lost in the conversion to MIDI from any other format, unless the music was written with the MIDI standards in mind, which is rare. Digital audio formats such as WAV, mp3, and ogg are used when the audio itself is part of the analysis. Lossy formats such as mp3 and ogg work well with the human ear but may be missing crucial data for study. Additionally some encodings create artifacts which could be misleading to any automatic analyser. Despite this the ubiquity of the mp3 has meant much research in the field involves these as the source material. Increasingly, metadata mined from the web is incorporated in MIR for a more rounded understanding of the music within its cultural context, and this recently consists of analysis of social tags for music.

Feature representation

Analysis can often require some summarising, and for music this is achieved by feature extraction, especially when the audio content itself is analysed and machine learning is to be applied. The purpose is to reduce the sheer quantity of data down to a manageable set of values so that learning can be performed within a reasonable time-frame. One common feature extracted is the Mel-Frequency Cepstral Coefficient which is a measure of the timbre of a piece of music. Other features may be employed to represent the key, chords, harmonies, melody, main pitch, beats per minute or rhythm in the piece. There are a number of available audio feature extraction tools

Statistics and machine learning