Speech signals are studied with a new method of time series analysis w
hich is stable with respect to nonlinear distortion. Separate phonemes
are characterized by continuous functions in the unit square. Possibl
e applications to speech and speaker recognition are indicated.