A Non-Linear Operator based Method for Harmonic Feature Extraction from Speech Signals
Darren F. Kavanagh, Frank Boland · 2007
An important pre-processing stage in speech recognition systems is that of extracting phonetically pertinent acoustic features from the speech signal. These features form the basis for discriminative classification and serve as cues for the identification of phonetic events in speech. The paper addresses this by presenting a novel method for the classification of harmonic (short-term periodic) and non-harmonic segments in speech signals. Classification is accomplished by proposing two new features derived from the non-linear Teager energy operator (TEO). The features proposed are the TEO-Weighted Harmonic Product (TEO-WHP*) and the TEO-Weighted Harmonic Sum (TEO-WHS*). Experiments are reported and discussed that demonstrate the effectiveness and the importance of these features as a valuable pre-processor for many speech systems.