CITATION — REFERENCE ENTRY

Deep clustering: Discriminative embeddings for segmentation and separation — arXiv

Revision a0dd93fc-cd57-4b49-b1b2-479fe556d559 · 11/21/2026, 6:45:54 AM UTC
Key
hershey2016-deep-clustering
Authors
Hershey, John R.; Chen, Zhuo; Le Roux, Jonathan; Watanabe, Shinji
Issued
2015-8-18
Type
article
Container
arXiv
Raw CSL JSON
{
  "URL": "https://arxiv.org/abs/1508.04306",
  "note": "Claims verified from the arXiv preprint (v1). A shorter version was published in Proceedings of ICASSP 2016, pp. 31-35, doi:10.1109/ICASSP.2016.7471631.",
  "type": "article",
  "title": "Deep clustering: Discriminative embeddings for segmentation and separation",
  "author": [
    {
      "given": "John R.",
      "family": "Hershey"
    },
    {
      "given": "Zhuo",
      "family": "Chen"
    },
    {
      "given": "Jonathan",
      "family": "Le Roux"
    },
    {
      "given": "Shinji",
      "family": "Watanabe"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2015,
        8,
        18
      ]
    ]
  },
  "number": "1508.04306",
  "container-title": "arXiv"
}

Claims

  1. Applying perceptual grouping theory to audio segmentation is generally known as computational auditory scene analysis.
    "Application of the perceptual grouping theory to audio segmentation is generally known as computational auditory scene analysis (CASA) [14, 15]."
    Locator: page: 2 (sec. 1) · Quote language: en
  2. Rather than estimating signals or masks directly, deep clustering trains a deep network to produce spectrogram embeddings that are discriminative for partition labels.
    "Rather than directly estimating signals or masking functions, we train a deep network to produce spectrogram embeddings that are discriminative for partition labels given in training data."
    Locator: page: 1 (Abstract) · Quote language: en
  3. The training objective is designed to be invariant to the number and permutations of partition labels.
    "we need a convenient expression that is invariant to the number and permutations of the partition labels from one training example to the next."
    Locator: page: 3 (sec. 2) · Quote language: en
  4. Trained on two-speaker mixtures and tested on held-out speakers, deep clustering improved signal quality by around 6 dB and generalized to three-speaker mixtures.
    "when trained on spectrogram features containing mixtures of two speakers, and tested on mixtures of a held-out set of speakers, it can infer masking functions that improve signal quality by around 6dB. We show that the model can generalize to three-speaker mixtures despite training only on two-speaker mixtures."
    Locator: page: 1 (Abstract) · Quote language: en
Available in