Reduction of LPC Spectrum Dimension Using a Wine-Glass-Type Neural Network

4 References

Bibliographic Information

Other Title
  • 砂時計型ニューラルネットによるLPCスペクトルの次元圧縮

Search this article

Description

Reducing the dimension of acoustic feature space is realized using a wine-glass-type neural network, which has the fewer number units in middle layer than the input and output layers, trained for the identity mapping. A wine-glass-type neural network, which has 32 units for both input and output layers and two to five units for the middle layer are trained so as to map the input of 32 dimensional LPC spectrum to the identical output vectors. After neural network is trained, signal to deviation ratio (SDR) of log spectrum is smaller than using KL expansion. Moreover, DTW isolated word recognition experiments are performed using 123 similar city name utterances of a male speaker. Using the output of the middle layer units reduced to 3-5 feature vector, the recognition accuracy are higher than using KL expansion. Therefore the effectiveness of nonlinear identity mapping using neural network for reducing the feature dimension is confirmed.

Journal

  • Technical report of IEICE. DSP

    Technical report of IEICE. DSP 96 (239), 39-44, 1996-09-13

    The Institute of Electronics, Information and Communication Engineers

References(4)*help

See more

Keywords

Details 詳細情報について

  • CRID
    1573387452252465536
  • NII Article ID
    110003279652
  • NII Book ID
    AN10060786
  • ISSN
    09135685
  • Text Lang
    ja
  • Data Source
    • CiNii Articles

Report a problem

Back to top