Automatic Music Transcription is the process to extract an acceptable music notation from performed data. One of the main problems in this area is the quantization and categorization of note durations. This is because in music performance large deviations from the notated values occur, such as playing notes earlier or later to accent them and by global tempo variations (e.g., slowing down at the end of the piece). These timing variations are musically meaningful and intended by the performer. They are linked directly to musical structure (e.g., the meter, phrase and rhythmical structure). Quantization can thus be defined as separating the categorical, discrete timing component (as notated in the score), from the continuous deviations.

Earlier approaches, such as the Stanford music transcription project, reached some level of success by incorporating large quantities of musical knowledge in the system. However, this symbolic, rule-based approach became very style specific (i.e. domaindependence) and broke easily when applied outside that type of music (i.e. brittleness). In our own work we have been able to show that a connectionist approach suffers less from these drawbacks. A network of cells representing all implicit time intervals in the input pattern, with an interaction which adjusts cells representing neighboring intervals towards integer multiples can quickly discover the rhythmic regularity in the input: it relaxes into a state representing the note duration categories. This connectionist quantizer does not need specific style knowledge and shows graceful degradation (an error might occur when complicated rhythms are performed, but it is always a simpler version of the pattern, not an absurd one).

We propose to elaborate this algorithm and move it towards a robust component for music transcription systems. We will also embed the method closely into the field of artificial neural networks to make the methods and theoretical results applicable to this kind of net. Methods for learning will be added to the system and the relaxation strategy will be improved. Furthermore, the original algorithm needs to be extended with a tempo track component (to allow the transcription of long musical fragments) and processing of polyphonic data.

Resultaten van het onderzoek