CITATION — REFERENCE ENTRY
On training targets for supervised speech separation — IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Key
- wang2014-training-targets
- Authors
- Wang, Yuxuan; Narayanan, Arun; Wang, DeLiang
- Issued
- 2014-12
- Type
- article-journal
- Container
- IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Volume
- 22
- Issue
- 12
- Pages
- 1849-1858
Raw CSL JSON
{
"DOI": "10.1109/TASLP.2014.2352935",
"URL": "https://pmc.ncbi.nlm.nih.gov/articles/PMC4293540/",
"PMID": "25599083",
"page": "1849-1858",
"type": "article-journal",
"PMCID": "PMC4293540",
"issue": "12",
"title": "On training targets for supervised speech separation",
"author": [
{
"given": "Yuxuan",
"family": "Wang"
},
{
"given": "Arun",
"family": "Narayanan"
},
{
"given": "DeLiang",
"family": "Wang"
}
],
"issued": {
"date-parts": [
[
2014,
12
]
]
},
"volume": "22",
"container-title": "IEEE/ACM Transactions on Audio, Speech, and Language Processing"
}
Claims
-
The ideal binary mask has traditionally been the training target for supervised speech separation because of its simplicity and large intelligibility gains.
"Traditionally, the ideal binary mask (IBM) is used as the target because of its simplicity and large speech intelligibility gains."
-
The ideal ratio mask, with a tunable exponent, is closely related to the frequency-domain Wiener filter.
"β is a tunable parameter to scale the mask. Although technically different, one can see that the IRM is closely related to the frequency-domain Wiener filter assuming speech and noise are uncorrelated"
-
Masks make direct contact with the observed mixtures, since they are used to modulate the mixtures in the time-frequency domain.
"Perhaps most importantly, masks make direct contact with the observed mixtures in the sense that they are used to modulate the mixtures in the time-frequency domain."
-
The two ratio-mask targets, the IRM and the FFT-MASK, outperformed the other targets on objective intelligibility and quality metrics.
"Our results in various test conditions reveal that the two ratio mask targets, the IRM and the FFT-MASK, outperform the other targets in terms of objective intelligibility and quality metrics."
Available in