Novel features for robust speech recognition
Alexandros Potamianos · The Journal of the Acoustical Society of America · 2002
Recently there has been much research in the area of robust front-ends and new features for automatic speech recongition (ASR). These efforts have had limited success for certain databases and recording conditions. In this work, we review some recent work on features for ASR: the articulatory front-end of Li et al. (2000), nonlinear modulation features [Quatieri (2002); Dimitriadis and Maragos (2002)], chaotic features [Pitsikalis and Maragos (2002)], short-time spectral moments [Paliwal et al. (2000)], etc. We extend the work of Potamianos and Maragos (2001) to show how some of these features relate to the standard front-end of short-time smooth spectral envelope. We also analyze some of these new features using classification and regression trees to show the relevance of the features for phone-classification task.