Non-stationary noise estimation utilizing harmonic structure for spectral subtraction

The spectral subtraction method is well known as a speech enhancement technique. In this paper we study the noise spectrum estimation problem required for the SS method. It is set out to estimate directly the noise spectrum from the noise-corrupted speech frame. To accomplish this, a fundamental frequency is detected and harmonic structure of the speech signal is utilized. We assume that non-stationary noise consists of its stationary part and non-stationary part. The non-stationary part is estimated from the harmonics obtained, while the stationary part is estimated in the conventional way. Both parts are combined, resulting in an accurate estimate of the noise spectrum. It is shown by experiments that the proposed method provides an improvement of the performance of the SS method in non-stationary noise environments.