A novel spectral subtraction technique for cognitive radios

Abstract

In the past a lot of work has been done to remove noise from speech. Most of the presented techniques were derived from Boll's spectral subtraction technique. Roughly speaking the spectral subtraction techniques consists of estimating the noise power during the periods when no speech is present and subtracting this estimate of the noise power from the signal when speech is present. This spectral subtraction technique could be a very good in-band de-noising technique for communication signals measured by cognitive radios. However, there is one major drawback: one can never turn off the spectrum so that no communication signals are present. This paper presents an extended version of the spectral subtraction technique which does not require `speech free' periods, but can determine the noise power from the empty frequency bins in the spectrum. The presented method is based on an autoregressive (AR) model, which is linear in the parameters. Simulation results show that the presented technique is as performing as the original spectral subtraction techniques without the need to turn off the signals.