Robust estimation methods for impulsive noise suppression in speech

We discuss a new robust time domain filtering method that detects and reconstructs speech segments corrupted by impulsive noise. Robust statistical methods are very effective in the case of impulsive environments such as wireless communications and cellular phone applications. The speech signal may be corrupted by impulsive noise lasting several milliseconds. We utilize a robust estimator of covariance based on one-dimensional projections and sample median calculations to detect these impulsive segments. This method, called projection statistics, is a very computationally efficient algorithm to suppress the impulses. We estimate the missing segments of speech using the linear prediction technique whose parameters are estimated using a robust Schweppe-type Huber generalized maximum likelihood (GM) estimator. A robust estimator is needed since speech signals closely follow the Laplacian distribution rather than the Gaussian and edges from the impulses may be leftover in the signal. We provide preliminary simulation results from actual speech containing co-channel and fading interferences from cellular phones.