Transcribing Bach Chorales Using Particle Swarm Optimisations

Abstract

This paper reports a novel application of particle swarm optimisation to polyphonic transcription task. The system transforms an input audio into activation strength of pitches in the desired range. This transformation begins with audio information in time-domain to frequency-domain and finally, to activation strength of pitches (a.k.a. piano-roll representation). We can infer the likely sounding pitches by comparing the observed activation strength of input audio to reference Tone-models. Although each Tone-model is learned offline from the pitches one wish to perform transcription with, this process often only approximates the Tone-model characteristics due to the variations in volume and other effects introduced from the manner of note executions. Hence, predicting sounding notes based solely on Tone-models gives inaccurate predictions. Here, we apply PSO to search for an optimum aggregation of different predicted pitches that best represents the input activation strength. We describe our problem formulation and the design of our approach. The experimental results show our approach to be of potential in the task of polyphonic transcription.