Study on the Effect of Feedback Loops and Cross-Coupled Hidden Layers in NN for Pitch Extraction

MIYABAYASHI Hideo, FUNADA Tetsuo

Bibliographic Information

Other Title

ピッチ抽出NNにおけるフィードバックおよび層内相互結合の効果の検討

Description

To extract pitch frequencies of speech signals is one of the most important subjects in speech analysis and synthesis. In this paper, we propose three multi-layered networks which have possibility of learning time-continuity and smooth fluctuation of pitch or U/V, and learning discriminant function with a high accuracy. The results of experiments using continuous speech samples show that a sequential NN with cross-coupled hidden layers has higher performance for pitch extraction and U/V detection than ordinary multi-layered NN, since state layers in feedback loops and cross-coupled hidden layers can effectively learn time-continuity of U/V or smoothing of pitch fluctuation, and can improve generalization ability of discrimination learning.

Journal

IEICE technical report. Speech

IEICE technical report. Speech 95 (140), 31-38, 1995-07-13

The Institute of Electronics, Information and Communication Engineers

Keywords

Details 詳細情報について

CRID: 1570572702385717632

NII Article ID: 110003296488

NII Book ID: AN10013221

Text Lang: ja

Data Source

CiNii Articles

Export

Report a problem