FEATURE EXTRACTION FROM NOISY SPEECH SIGNALS
The accurate extraction of three characteristic features of a speech signal contaminated by high levels of additive white gaussian noise is investigated. Reliable determination of the formant frequencies provides important information on the characteristics of the vocal tract transfer function. Accurate measurement of the fundamental frequency and voiced/unvoiced categorization of a speech signal provide important information on the characteristics of the excitation of the vocal tract. These three features are of high importance in many speech processing systems. A highly noise immune, zero-crossings consistency phenomenon is utilized in this work for accurate measurement of these three important features from speech contaminated by high levels of additive white gaussian noise. Extensive testing has demonstrated excellent accuracy in extracting these three features for speech contaminated by high levels of additive white gaussian noise.
LAHAT, MEIR, "FEATURE EXTRACTION FROM NOISY SPEECH SIGNALS" (1983). Dissertations (1962 - 2010) Access via Proquest Digital Dissertations. AAI8317278.