A Bassline Generation System Based on Sequence-to-Sequence Learning

Abstract

This thesis presents a detailed explanation of a system generating basslines that are stylistically and rhythmically interlocked with a provided audio drum loop. The proposed system is based on a natural language processing technique: wordbased sequence-to-sequence learning. The word-based sequence-to-sequence learning method proposed in this thesis is comprised of recurrent neural networks composed of LSTM units. The novelty of the proposed method lies in the fact that the system is not reliant on a voice-by-voice transcription of drums; instead, in this method, a drum representation is used as an input sequence from which a translated bassline is obtained at the output. The drum representation consists of fixed size sequences of onsets detected from a 2-bar audio drum loop in eight different frequency bands. The basslines generated by this method consist of pitched notes with different duration. The proposed system was trained on two distinct datasets compiled for this project by the authors. Each dataset contains a variety of 2-bar drum loops with annotated basslines from two different styles of dance music: House and Soca. A listening experiment designed based on the system revealed that the proposed system is capable of generating basslines that are interesting and are well rhythmically interlocked with the drum loops from which they were generated.