H. Marvi - Department of Electrical Engineering Shaharood Universitiy of TechnologyShahrood/Iran
E. Chilton - Centre for Vision, Speech and Signal ProcessingUniversity of Surrey, Guildford, UK

It has been shown that detailed information from non-stationary signals such as speech are better represented by an acoustic image, a two dimensional feature representation [10]. Several time frequency representations such as the spectrogram, Wigner-ville and choi-williams distribution have been proposed [6] while the acoustic images based on the two dimensional root cepstrum analysis (TDRC) is a special case [5]. The novel distribution in this paper suggests acoustic images based on Hartley two dimensional root cepstrum (HTDRC) analysis to represent a non-stationary signals such as speech and preserve both magnitude and phase details of the signal simultaneously. Furthermore it has the capability to extract detailed of both static and dynamic features of the signal. Experimental results demonstrate that the acoustic images based on the HTDRC outperforms the TDRC in speech recognition applications. This method increases the recognition accuracy by 8.1%

Acoustic image, Speech Recognition, Feature extraction, Hartley transform