A New Algorithm for Speech Feature Extraction Using Polynomial Chirplet Transform

Time-frequency analysis (TFA) is a powerful tool for signal feature representation. In the time-frequency plane, the primary data properties are shown with both instantaneous values and trends of frequency change during time. With a complicated and non-stationary signal such as human speech, the conventional TFA tools, including Fourier transform, wavelet transform, or linear chirplet transform (LCT), cannot reveal and represent speech behaviors well. This research proposes a new method for speech representation with a TFA perspective using polynomial chirplet transform (PCT). Inspired by the Weierstrass theorem, PCT uses a polynomial function for instantaneous frequency (IF) estimation. This polynomial also shapes the modulated atom for the transform. With the strength of a high-degree polynomial, PCT can capture many meaningful features in human speech and then robust the recognition models by improving the features representation. Experimental results in the speech processing tasks have demonstrated the potential of PCT. Furthermore, it will perform better if PCT is optimized with an adaptive strategy to identify the IF function.