Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Voice Parameter Estimation Using Sequential SVD and Wave Shaping Filter Bank

Sunghoon Hong, Sangki Kang, Souguil Ann

Department of Electronics Engineering, Seoul National University, Seoul, Korea

In this paper, we propose new estimation technique of voice source and vocal-tract parameters in voiced speech. First, Kalman filtering technique which is a new approach to time-varying analysis is examined. Second, in order to overcome the problems of Kalman filtering approach, the error covariance matrix in Kalman filter is decomposed by reduced-rank SVD(singular value decomposition). Third, in order to estimate more accurate vocal-tract parameters we introduce robustness concept which assumes residual characteristics as normal mixture density. And, finally we propose wave-shaping filter bank approach as a new voice source modeling method which can represents and controls various source characteristics in frequency domain. The proposed methods give more accurate and more various time-varying voice parameters. These methods can be used to improve the speech quality in speech synthesis and coding techniques.

Full Paper

Bibliographic reference.  Hong, Sunghoon / Kang, Sangki / Ann, Souguil (1994): "Voice parameter estimation using sequential SVD and wave shaping filter bank", In ICSLP-1994, 1059-1062.