CITATION — REFERENCE ENTRY

Monaural speech segregation based on pitch tracking and amplitude modulation — IEEE Transactions on Neural Networks

Revision 67c1dcdb-c0d6-4e8b-bfcf-efe4200a1889 · 11/21/2026, 6:45:51 AM UTC
Key
hu2004-pitch-am
Authors
Hu, Guoning; Wang, DeLiang
Issued
2004-9
Type
article-journal
Container
IEEE Transactions on Neural Networks
Volume
15
Issue
5
Pages
1135-1150
Raw CSL JSON
{
  "DOI": "10.1109/TNN.2004.832812",
  "URL": "https://pnlwang.github.io/papers/Hu-Wang.tnn04.pdf",
  "page": "1135-1150",
  "type": "article-journal",
  "issue": "5",
  "title": "Monaural speech segregation based on pitch tracking and amplitude modulation",
  "author": [
    {
      "given": "Guoning",
      "family": "Hu"
    },
    {
      "given": "DeLiang",
      "family": "Wang"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2004,
        9
      ]
    ]
  },
  "volume": "15",
  "container-title": "IEEE Transactions on Neural Networks"
}

Claims

  1. For unresolved harmonics, the system forms segments from common amplitude modulation and groups them by AM rate.
    "For unresolved harmonics, it generates segments based on common amplitude modulation (AM) in addition to temporal continuity and groups them according to AM rates."
    Locator: page: 1135 (Abstract) · Quote language: en
  2. Hu and Wang's system segregates voiced speech, treating resolved and unresolved harmonics differently.
    "We propose a novel system for voiced speech segregation that segregates resolved and unresolved harmonics differently."
    Locator: page: 1135 (Abstract) · Quote language: en
  3. Using the ideal binary mask as ground truth, the system retained on average 96.36% of target speech energy with 2.83% noise residue.
    "we here suggest to use an ideal binary mask as the ground truth of target stream. ... On average, our system retains 96.36% of target speech energy, and the percentage of noise residue is kept at 2.83%."
    Locator: page: 1146 (sec. VII) · Quote language: en
  4. Grouping only by pitch limits the model to voiced speech; the authors see unvoiced speech as the biggest challenge for monaural segregation.
    "Our model performs grouping based only on pitch. As a result, it is limited to segregation of only voiced speech. In our view, unvoiced speech poses the biggest challenge for monaural speech segregation."
    Locator: page: 1149 (sec. VIII) · Quote language: en
Available in