CITATION — REFERENCE ENTRY

Permutation invariant training of deep models for speaker-independent multi-talker speech separation — 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

Revision 356fd9e7-fb6d-4a51-812a-32b98f51c291 · 11/21/2026, 6:45:54 AM UTC
Key
yu2017-pit
Authors
Yu, Dong; Kolbæk, Morten; Tan, Zheng-Hua; Jensen, Jesper
Issued
2017-3
Type
paper-conference
Container
2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
Pages
241-245
Raw CSL JSON
{
  "DOI": "10.1109/ICASSP.2017.7952154",
  "URL": "https://arxiv.org/abs/1607.00325",
  "note": "Claims verified from arXiv 1607.00325 v2; page locators refer to that version.",
  "page": "241-245",
  "type": "paper-conference",
  "title": "Permutation invariant training of deep models for speaker-independent multi-talker speech separation",
  "author": [
    {
      "given": "Dong",
      "family": "Yu"
    },
    {
      "given": "Morten",
      "family": "Kolbæk"
    },
    {
      "given": "Zheng-Hua",
      "family": "Tan"
    },
    {
      "given": "Jesper",
      "family": "Jensen"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2017,
        3
      ]
    ]
  },
  "container-title": "2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)"
}

Claims

  1. Before deep learning, CASA was the most popular technique for multi-talker separation, using often semi-manually designed segmentation rules based on perceptual grouping cues to estimate a time-frequency mask.
    "Before the deep learning era, the most popular technique was computational auditory scene analysis (CASA) [9, 10]. In this approach, certain segmentation rules based on perceptual grouping cues [11] are (often semi-manually) designed to operate on low-level features to estimate a time-frequency mask"
    Locator: page: 1 (sec. 1) · Quote language: en
  2. Yu and colleagues judge that CASA and NMF had very limited success in separating multi-talker speech.
    "Both CASA and NMF led to very limited success in separating sources in multi-talker mixed speech [7]."
    Locator: page: 1 (sec. 1) · Quote language: en
  3. Permutation invariant training computes the error for each assignment of outputs to references and optimizes the model on the assignment with the least error.
    "The assignment with the least total MSE is chosen and the model is optimized to reduce this particular MSE."
    Locator: page: 3 (sec. 3) · Quote language: en
  4. The label ambiguity, or permutation, problem caused earlier methods to perform poorly on speaker-independent multi-talker separation.
    "This problem is referred to as the label ambiguity (or permutation) problem in [8, 23]. Due to this problem, prior arts perform poorly on speaker-independent multi-talker speech separation."
    Locator: page: 2 (sec. 3) · Quote language: en
Available in