Presentation 1995/5/26
Compression of Speech Spectrum Dimension Using a Wine-Glass-Type Neural Network
Hironori Ito, Fumitada Itakura,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In this paper, speech spectrum of 32 dimensions are compressed to 2-4 dimensions using a wine-glass-type neural network to implement an identity mapping. Then, the mean spectral distortion for each number of dimensions is examined. Two types of neural network(namely three and five layers) are used. It is observed that the spectral distortion is much smaller for the five-layers neural network than for the three-layers. It is also verified that a pre-devided-lerning is effective for the five-layers neural network. The distribution of middle layer's outputs of the neural network is examined. For the case of vowel data, a cluster distribution for each vowel is well separated from each others.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) wine-glass-type neural network / speech spectrum / identitiy mapping / dimension compression / pre-divided-learning
Paper #
Date of Issue

Conference Information
Committee EA
Conference Date 1995/5/26(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To Engineering Acoustics (EA)
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Compression of Speech Spectrum Dimension Using a Wine-Glass-Type Neural Network
Sub Title (in English)
Keyword(1) wine-glass-type neural network
Keyword(2) speech spectrum
Keyword(3) identitiy mapping
Keyword(4) dimension compression
Keyword(5) pre-divided-learning
1st Author's Name Hironori Ito
1st Author's Affiliation Graduate School of Engineering, Nagoya University()
2nd Author's Name Fumitada Itakura
2nd Author's Affiliation Graduate School of Engineering, Nagoya University
Date 1995/5/26
Paper #
Volume (vol) vol.95
Number (no) 70
Page pp.pp.-
#Pages 8
Date of Issue