CITATION — REFERENCE ENTRY

SingSong: Generating musical accompaniments from singing — arXiv

Revision b63d72dc-cb55-4786-9ae3-aa070c33e773 · 11/1/2026, 5:06:40 PM UTC
Key
donahue2023-singsong
Authors
Donahue, Chris; Caillon, Antoine; Roberts, Adam; Manilow, Ethan; Esling, Philippe; Agostinelli, Andrea; Verzetti, Mauro; Simon, Ian; Pietquin, Olivier; Zeghidour, Neil; Engel, Jesse
Issued
2023-1-30
Type
article
Container
arXiv
Raw CSL JSON
{
  "URL": "https://arxiv.org/abs/2301.12662",
  "type": "article",
  "title": "SingSong: Generating musical accompaniments from singing",
  "author": [
    {
      "given": "Chris",
      "family": "Donahue"
    },
    {
      "given": "Antoine",
      "family": "Caillon"
    },
    {
      "given": "Adam",
      "family": "Roberts"
    },
    {
      "given": "Ethan",
      "family": "Manilow"
    },
    {
      "given": "Philippe",
      "family": "Esling"
    },
    {
      "given": "Andrea",
      "family": "Agostinelli"
    },
    {
      "given": "Mauro",
      "family": "Verzetti"
    },
    {
      "given": "Ian",
      "family": "Simon"
    },
    {
      "given": "Olivier",
      "family": "Pietquin"
    },
    {
      "given": "Neil",
      "family": "Zeghidour"
    },
    {
      "given": "Jesse",
      "family": "Engel"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2023,
        1,
        30
      ]
    ]
  },
  "number": "2301.12662",
  "container-title": "arXiv"
}

Claims

  1. SingSong generates instrumental music audio to accompany input vocals. It was trained on vocal and instrumental pairs obtained by applying source separation to a large corpus of music, using an adaptation of the AudioLM audio generation model; listeners significantly preferred its instrumentals to those from a retrieval baseline.
    "We present SingSong, a system that generates instrumental music to accompany input vocals [...]. Specifically, we apply a state-of-the-art source separation algorithm to a large corpus of music audio to produce aligned pairs of vocals and instrumental sources. Then, we adapt AudioLM [...] to be suitable for conditional "audio-to-audio" generation tasks, and train it on the source-separated (vocal, instrumental) pairs. In a pairwise comparison with the same vocal inputs, listeners expressed a significant preference for instrumentals generated by SingSong compared to those from a strong retrieval baseline."
    Locator: section: Abstract · Quote language: en
Available in