Genetic Algorithm Approach to Polyphonic Music Transcription

Abstract

This paper describes our method, submitted to MIREX 2008 task "Multiple Fundamental Frequency Estimation & Tracking". This task restricted the problem of Multiple F0 Estimation and Tracking to three cases: i) Estimate active fundamental frequencies on a frame-by-frame basis; ii) Track note contours on a continuous time basis (as in audio-tomidi); iii) Track timbre on a continuous time basis. The presented method is the first genetic algorithm approach to multi-timbral music transcription and is based on a previously published genetic algorithm approach for automatic music transcription originally designed to transcribe piano songs. This algorithm was adopted to support different timbres.

Automatic Transcription of Polyphonic Piano Music Using Genetic Algorithms, Adaptive Spectral Envelope Modeling, and Dynamic Noise Level Estimation

“…According to our previous results [19], the initial population has two major contributions to the end result. First, if the initial population is created nearer (or even half way) to the final result than a randomly generated initial population, the genetic algorithm will need a much smaller number of generations to achieve the target result.…”

“…Also, it cannot detect note dynamics. Later in 2007, Reis et al [19] proposed the first genetic algorithm approach to music transcription using real audio recordings. The authors explored the influence of the "harmonic overfitting" phenomena, which is related to differences in the spectral envelope between different pianos.…”

Toward Evolution Strategies Application in Automatic Polyphonic Music Transcription using Electronic Synthesis

“…However this approach suffers from its computationally intensive behaviour and the convergence time which is very long due to the application of GAs in the context of polyphonic music transcription. In [5], GAs was also used for polyphonic music transcription and moreover, they were applied to address multi-timbral music transcription.…”

“…Gustavo Reis et al also used GAs for automatic polyphonic music transcription by addressing the multiple fundamental frequencies problem and tracking [14]. Their approach was restricted to three cases which are: estimation of active fundamental frequencies on a frame-by-frame basis; tracking of note contours on a continuous time basis; tracking of timbre on a continuous time basis.…”

“…These latter ones will enable the ES strategy to optimize only the related segment of the audio signal representing a sound played at a certain interval of time. Therefore, improving the accuracy of the results compared to others methods, which need to encode the full audio signal in a chromosome [5], [6]. At the end of the execution, the results from all ES which took as input each audio segment are returned in the correct order along with their durations.…”