A modified frequency-weighted Itakura spectral distortion measure

A modified frequency-weighted Itakura spectral distortion measure (FWIDM) for speech recognition in noise is described. It is shown that if noise is added to the test speech signal before preemphasis (such as would be the case with environment noise picked up at the microphone), the FWIDM method is not effective in reducing the distance between the noisy test spectrum and the clean reference spectrum. The modified FWIDM overcomes this problem. The modification is to use the bandwidth broadened spectrum of the nonpreemphasized test speech signal as the weighting function. It is shown that this weighting function reduces the distance between the clean reference spectrum and the noisy test spectrum of the preemphasized speech signals. This weighting function is also effective if the test and reference spectra are obtained from the nonpreemphasized speech signal