Iterative spectral subtraction based on variable frame processing for musical noise reduction

In recent years, speech communication and speech applications like speech recognition are widely used with the development of information equipment. But a speech corrupted by noises makes fluent communication difficult and causes the poor performance of speech applications. Thus, an accurate noise reduction method is necessary to solve these problems. Spectral subtraction (SS) and iterative spectral subtraction (I-SS) are well-known noise reduction methods that subtract estimated noise from noisy speech in the power spectral domain. However, SS and I-SS generate an artificial noise called musical noise. The noise is non-stationary and has an offensive sound. Miyazaki et al. have proposed a musical-noise-free SS that determines subtraction parameters such that musical noise is not generated. The method effectively reduces the musical noise but requires estimation of the statistical characteristics of the observed noise. Meanwhile, we tackle the musical noise problem by proposing a new I-SS based on variabl...