Comparing Models for Harmony Prediction in an Interactive Audio Looper

SongDriver: Real-time Music Accompaniment Generation without Logical Latency nor Exposure Bias

“…Latency Models Bias Models SongDriver PSCA-HMM [24] PSCA-RNN [24] Music Transformer [11] Markov-Lin [14] current melody more harmoniously compared with other latency models, which corroborates that the generated accompaniment of SongDriver is more logically synchronized with the melody. Besides, the CS score also justifies the high harmonic stability of SongDriver generated accompaniments, proving the effectiveness of the twophase strategy.…”

“…These existing models could be classified into two categories: latency models and bias models. Latency models can be represented by [24], which uses HMM and RNN to ensure the accuracy of the improvised accompaniment. But the accompaniment often slightly lags behind its corresponding melody, thus leading to a logical latency.…”

“…Previous automatic real-time accompaniment models are generally one-phase, and can be divided into two categories on the basis of the said two problems: 1) Latency models [23] [24], which avoid the exposure bias but suffer from the logical latency by arranging the accompaniment each time the input reaches a preset length. 2) Bias models [1] [12], which eliminate the logical latency but face the exposure bias by generating accompaniments for the upcoming melody.…”