An Evaluation of the Harmonic Product Spectrum for Neural Network-Based Chord Recognition

The low-power computing systems have come under great demand today. This is due to the increasing energy consumption of connected devices and digital infrastructure. In the domain of chord recognition, there is a challenge to find feature representation methods that are computationally low while still preserving a high level of accuracy. In this study, the effectiveness of the Harmonic Product Spectrum (HPS) as a feature representation method for neural network chord recognition is evaluated. This chord recognition can be targeted for a small-scale and low-power system. Experiments were carried out using eight different HPS levels, where increasing the HPS level corresponded to a proportional reduction in the input size of the neural network. Based on the experimental results, it was shown that using HPS level 7, the chord recognition system could achieve an accuracy of up to 97.14%. These results indicate that HPS level 7 can provide the optimal trade-off between computational efficiency and accuracy.