Centre for

Speech Enhancement Tutorial - Enhancement Methods

3.9 Harmonic and Impulsive Noise

In this section we consider noise sources that are specifically bounded in the time-frequency domain. Interference that is temporally short can generally be considered impulsive, noises that are localized in frequency (whether individually or as a set of related components) can generally be considered harmonic. It is often possible to identify and clean these types of noise because they are well localized and the majority of the time-frequency representation of the signal is undamaged. Removing such interference can improve intelligibility because temporal and spectral masking in human hearing means that nearby regions of the time frequency plane will be inaudible despite being themselves uncorrupted.

Harmonic noise sources are usually heard as hums and buzzes which exhibit strong periodic components and corresponding harmonic structure. Suppression of harmonic noise that is temporally stationary usually begins with identifying the fundamental frequency. This can be achieved by manual spectral measurement or, in principle, by automatic methods. Notch filtering can then be applied both to the fundamental and to its harmonics to achieve suppression. When harmonic noise is time varying, it is necessary to track the fundamental and a more sophisticated approach, such as single channel adaptive filtering is required.

Examples of impulsive noise are most commonly clicks, pops and crackles. Such noise may result from a variety of sources including impulsive radio frequency interference, faulty electrical connections and digital recording errors. Methods for identifying and cleaning impulsive noise are extensively discussed in Godsill and Rayner [1998] and the references therein. The common techniques usually require manual localisation of the impulsive noise. With knowledge of the time of occurrence of impulse noise, methods can then be used to interpolate the undamaged signal that exists before and after the noise in order to remove it. Pure interpolation methods assume that no useful information remains at the instant of the impulsive noise.

Other methods assume that the original data may still be present during the impulsive noise and attempt to model it in order to achieve better accuracy. The statistical distribution of the samples, whether known, assumed or deduced, can be used to condition the interpolation processing. Autoregressive interpolation assumes the data can be well modelled by an autoregressive model and estimates the parameters of the autoregressive model using, for example, least squares error minimisation. Pitch information may also be employed to constrain the interpolated speech further.

Previous | Next