An Audio-to-Audio Approach to Generate Bass Lines from Guitar's Chord Backing

In this paper, we address a system that generates a bass line from a chord accompaniment played on the guitar based on an audio-to-audio manner. Yielding bass lines for guitar's chord backings would be useful for amateur musicians composing band music. Conventional music arrangement systems targeted MIDI-like symbolic music representations, but accurately obtaining symbolic representations from guitars is not easy. To solve this problem, we attempt an audio-to-audio approach; Once the user gives an audio recording of the guitar's chord backing, the system extracts some audio features (spectrogram, mel-spectrogram, or chromagram) and then generates an audio signal of bass lines using a convolutional neural network. The experimental results showed that the model with chromagrams generates bass lines the most robustly.