CITATION — REFERENCE ENTRY
Adaptive accompaniment with ReaLchords — Proceedings of the 41st International Conference on Machine Learning
- Key
- wu2024-realchords
- Authors
- Wu, Yusong; Cooijmans, Tim; Kastner, Kyle; Roberts, Adam; Simon, Ian; Scarlatos, Alexander; Donahue, Chris; Tarakajian, Cassie; Omidshafiei, Shayegan; Courville, Aaron; Castro, Pablo Samuel; Jaques, Natasha; Huang, Cheng-Zhi Anna
- Issued
- 2024
- Type
- paper-conference
- Container
- Proceedings of the 41st International Conference on Machine Learning
- Volume
- 235
- Pages
- 53328-53345
Raw CSL JSON
{
"URL": "https://proceedings.mlr.press/v235/wu24c.html",
"page": "53328-53345",
"type": "paper-conference",
"title": "Adaptive accompaniment with ReaLchords",
"author": [
{
"given": "Yusong",
"family": "Wu"
},
{
"given": "Tim",
"family": "Cooijmans"
},
{
"given": "Kyle",
"family": "Kastner"
},
{
"given": "Adam",
"family": "Roberts"
},
{
"given": "Ian",
"family": "Simon"
},
{
"given": "Alexander",
"family": "Scarlatos"
},
{
"given": "Chris",
"family": "Donahue"
},
{
"given": "Cassie",
"family": "Tarakajian"
},
{
"given": "Shayegan",
"family": "Omidshafiei"
},
{
"given": "Aaron",
"family": "Courville"
},
{
"given": "Pablo Samuel",
"family": "Castro"
},
{
"given": "Natasha",
"family": "Jaques"
},
{
"given": "Cheng-Zhi Anna",
"family": "Huang"
}
],
"issued": {
"date-parts": [
[
2024
]
]
},
"volume": "235",
"event-place": "Vienna, Austria",
"container-title": "Proceedings of the 41st International Conference on Machine Learning",
"collection-title": "Proceedings of Machine Learning Research"
}
Claims
-
ReaLchords is an online generative model that improvises chord accompaniment to a user's melody, generating each chord without knowing the melody note it will accompany. Models trained only by maximum likelihood transfer poorly to live settings with mistakes and fail to adapt; ReaLchords is finetuned with reinforcement learning using self-supervised reward models and distillation from a teacher model that can see the future melody. It was trained on about 38,000 melody–chord pairs from the Hooktheory dataset and evaluated in a listening test with ten musicians.
"We propose ReaLchords, an online generative model for improvising chord accompaniment to user melody. We start with an online model pretrained by maximum likelihood, and use reinforcement learning to finetune the model for online use. The finetuning objective leverages both a novel reward model that provides feedback on both harmonic and temporal coherency between melody and chord, and a divergence term that implements a novel type of distillation from a teacher model that can see the future melody. [...] Each chord must be generated without knowing in advance which melody note it will accompany. [...] Likelihood models, however, suffer from exposure bias due to being trained entirely on ground truth data, and transfer poorly to the online settings where mistakes, imperfections and stylistic differences are common [...]. We train our models on an updated version of the Hooktheory dataset [...] now contains 38K melody-chord pairs. [...] We recruited ten musicians, and collected 192 ratings"
-
In ReaLchords' evaluation, an offline model that can see the future melody stayed closer to the test data than any online model: 63.73% of melody notes fell within the offline model's chords, versus 48.17–54.29% for the ReaLchords variants and 70.94% in the human-written test set.
"Note in Chord ↑, % [...] ReaLchords 48.17 ± 0.27 [...] ReaLchords-M 54.29 ± 1.55 [...] Offline MLE 63.73 [...] Test set 70.94 [...] The behavior of Offline MLE is closest to that of the test set, as is expected due to its ability to see the future input."
Available in