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

Abstract

This paper presents a new method for multiple fundamental frequency (F0) estimation on piano recordings. We propose a framework based on a genetic algorithm in order to analyze the overlapping overtones and search for the most likely F0 combination. The search process is aided by adaptive spectral envelope modeling and dynamic noise level estimation: while the noise is dynamically estimated, the spectral envelope of previously recorded piano samples (internal database) is adapted in order to best match the piano played on the input signals and aid the search process for the most likely combination of F0s. For comparison, several state-of-the-art algorithms were run across various musical pieces played by different pianos and then compared using three different metrics. The proposed algorithm ranked first place on Hybrid Decay/Sustain Score metric, which has better correlation with the human hearing perception and ranked second place on both onset-only and onset-offset metrics. A previous genetic algorithm approach is also included in the comparison to show how the proposed system brings significant improvements on both quality of the results and computing time.

A Comprehensive Review on Automatic Music Transcription: Survey of Transcription Techniques

“…Moreover, the considered count of chords was considerably larger in this approach. In 2012, Reis et al [29] presented a technique for multiple frequency (F0) evaluation on piano recording. Consequently, GA was exploited for analyzing the overlapping tones and also for searching the certain F0 combinations.…”

Piano Transcription in the Studio Using an Extensible Alternating Directions Framework

“…A large body of work involves elaborate methods exploiting the harmonicity of musical sounds to group detected spectral peaks into note objects [13]- [16]. Another group of methods employs probabilistic sinusoidal plus noise modeling, in which parameters such as onset and offset position, fundamental frequency, intensity and spectral envelope parameters are adjusted to match the observed signal using maximum a posteriori estimation [17] or genetic algorithms [18]. Another approach employs a hidden Markov model (HMM) [19], where each state corresponds to one possible combination of active notes, which requires elaborate heuristic state-space pruning strategies to be computationally feasible.…”

Computational models for rhythm and applications on human-machine interactions

“…Different candidates are combined, changed, replicated and rejected using evolutionary strategies, and after some iterations the desired transcription is obtained. Later, Reis et al [77] incorporated variance on the spectral envelope and a dynamic noise level analysis, significantly outperforming the previous approach [76].…”