Adaptive noise estimation using second generation and perceptual wavelet transforms

This paper describes the implementation and performance evaluation of three noise estimation algorithms using two different signal decomposition methods: a second-generation wavelet transform and a perceptual wavelet packet transform. These algorithms, which do not require the use of a speech activity detector or signal statistics learning histograms, are: a smoothing-based adaptive technique, a minimum variance tracking-based technique and a quantile-based technique. The paper also proposes a new and robust noise estimation technique, which utilises a combination of the quantile-based and smoothing-based algorithms. The performance of the latter technique is then evaluated and compared to those of the above three noise estimation methods under various noise conditions. Reported results demonstrate that all four algorithms are capable of tracking both stationary and nonstationary noise adequately but with varying degree of accuracy.