User Project Details

CondEntPitchTS

Instantaneous information content extraction from musical pitch timeseries

other

Max-Planck Gesellschaft

Max-Planck-Institut für Dynamik und Selbstorganisation

Music theorists since Leonard Meyer have emphasized the importance of the interplay and balance between expectation and surprise in musical compositions. It would be desirable to quantify this interplay comparing different composers, musical genres, etc. on the basis of information theoretical concepts. Unfortunately the length of compositions is by far not sufficient to estimate e.g. entropies and redundancy of melodies, as the number of datapoints required to reliably compute probability distributions grows factorially with the possible pitch values and their permutations. To circumvent this limitation, some authors used recurrent neural networks (RNN) as a proxy for the probability distributions necessary to compute entropies. However, these attempts have been limited to the use of LSTM models, which are known to perform poorly for sequences longer than ~100 steps. Transformer architectures could greatly improve the range of analyzable pieces, as well as the overall results quality.