Linear predictive, eigenvalue oriented pitch-contour measurement for forensic voice identification

Presents a novel pitch-contour measurement scheme for highly noise-contaminated speech signals as found in the forensic voice-identification problem. The base of the method is the synthesis of the covariance function that arises from an idealized description of the well-known SIFT-algorithm. The retrieval of the involved sinusoidal frequencies is carried out by means of an AR-model. The issues of proper AR-algorithm choice and model-order selection are considered, thus leading to a order-adapting scheme that performs extremely robustly. The adaptation is based on a stability test which is embedded in the corresponding order-recursive algorithm. The robustness of the proposed technique is evaluated with a large amount of speech-data for different kinds and levels of distortions.<>