Classification of Guitar Chords Using Deep Neural Networks

Abstract

The pitch is a measurement of the frequency of a sound wave, which is measured in Hertz (Hz), the SI unit of frequency, equal to one cycle per second. Pitch tracking is often referred to as audio processing and alias challenging task, especially for polyphonic recordings such as guitar chords, that involves estimating the fundamental frequency of audio signals. In addition, pitch tracking is more relevant to speech and music information retrieval than other audio processing tasks. With the use of deep-learning techniques and a range of proposed network architectures, frame-wise transcription and multi-pitch estimation (MPE) has recently seen colossal improvements in their ability to detect pitches simultaneously in monophonic or polyphonic music recordings. Although there are some techniques and algorithms perform very well on average, there are still some instances in which they fail to yield the pitch estimation correctly. Motivated by previous studies. we propose a deep learning approach for multi-pitch estimation sing the GuitarSet dataset, a publicly available dataset of audio recordings of individual notes and chords with the collection of annotations played on six-string acoustic and electric guitars. The proposed a data-driven pitch-tracking algorithm based on deep convolutional neural networks operating directly on frequency-domain waveforms. In our work, we used the Fast Fourier Transform (FFT) to extract features from the raw audio signal and input them into the base model. Furthermore, we compressed the data with max-pooling and cropped the extracted features from the audio data. As a result, the preprocessed data produced a better result than the direct input from the raw audio signals.