Voice Cognition by Wavelets Image Processing
Isamu Senoo, S. Hayano, Yoshitaka Saito · Journal of the Visualization Society of Japan · 2000
As is well known, voice cognition system of the personal computers becomes one of the popular command input methodologies. However, it is still remaining problem, which requires a hard training in order to build up a reliable voice database. Most of the voice cognition methods are based on a simple correlation analysis between the input and database voices. This leads to a relatively low cognition rate when using a poor trained database system. To overcome this difficulty, we are now developing a new voice cognition methodology, which cognizes an input voice as a solution of an inverse problem. Our inverse methodology requires each of the characteristic patterns representing essential feature of the single pronunciation. In the present paper, we describe that the characteristic patterns representing essential feature of the single pronunciation are derived by means of the simple Lissajous's diagram after the threshold operations. Using the database containing the characteristic patterns, an ill-posed linear system of equations for an each pronunciation input is established. Simple least squares leads to a good approximate solution of this system. Thus, we have succeeded in cognizing the single pronunciation. Further, we have tried to compress the characteristic patterns by means of the discrete wavelet transform. As a result, it is revealed that the Daubechie's 2nd order base function makes it possible to reduce the database into 25% quantity.