Blind channel identification in speech using the Long-Term Average Speech Spectrum

Estimation of the magnitude response of an unknown channel in single-microphone speech signals is considered. It is shown how the Long-Term Average Speech Spectrum (LTASS) can be used to identify the unknown channel and a blind channel identification algorithm is developed based on that. Furthermore, an established approximate formula for LTASS is demonstrated to be a useful tool in the context. The algorithm is evaluated using a weighted spectral distortion measure using simulated, measured and real channels with various distinct spectral characteristics. It is demonstrated that the algorithm can identify accurately the magnitude spectrum of an unknown channel in noise-free conditions. We also show results for three different additive noises where estimation accuracy is reduced but the degradation varies largely, depending on the long-term spectral characteristics of the noise.