Speech recognition using Hilbert-Huang transform based features

Abstract

The Mel-frequency Cepstral Coefficients (MFCCs) are widely used for feature extraction in Automatic Speech Recognition (ASR) systems. MFCCs start by dividing the speech into windows and calculating the Fourier Transform (FT) of each window. The frequency resolution obtained using this scheme depends on the time width of the window. A small window would fail to provide a good frequency resolution, while a big window would fail to obtain a good time resolution. This phenomenon is explained by the way the FT defines frequency; it tries to map the signal to a set of predefined bases. In this work, we propose a speech feature extraction method, we will refer to as Mel Hilbert Frequency Cepstral Coefficients (MHFCCs). MHFCCs use the Hilbert-Huang Transform (HHT) instead of the windowing and the FT scheme used in MFCCs. The HHT is an adaptive time-frequency transform suitable for non-linear and non-stationary signals. It generates its bases from the signal itself. This enables it to obtain a high frequency resolution representation of the signal regardless of the duration of the time window used. Results show that MHFCCs outperform MFCCs in recognition accuracy for a small time window.