Comparing Models for Harmony Prediction in an Interactive Audio Looper

Benedikte Wallace, Charles P. Martin

Abstract: Musicians often use tools such as loop-pedals and multitrack recorders to assist in improvisation and songwriting, but these tools generally don’t proactively contribute aspects of the musical performance. In this work, we introduce an interactive audio looper that predicts a loop’s harmony, and constructs an accompaniment automatically using concatenative synthesis. The system uses a machine learning (ML) model for harmony prediction, that is, it generates a sequence of chords symbols for a given melody. We analyse the performance of two potential ML models for this task: a hidden Markov model (HMM) and a recurrent neural network (RNN) with bidirectional long short-term memory (BLSTM) cells. Our findings show that the RNN approach provides more accurate predictions and is more robust with respect to changes in the training data. We consider the impact of each model’s predictions in live performance and ask: “What is an accurate chord prediction anyway?”