On the influence of inharmonicities in model-based speech enhancement

In relation to speech enhancement, we study the influence of modifying the harmonic signal model for voiced speech to include small perturbations in the frequencies of the harmonics. A perturbed signal model is incorporated in the nonlinear least squares method, the Capon filter and the amplitude and phase estimation filter. Results show that it is possible to increase the performance, in terms of the signal reduction factor and the output signal-to-noise ratio, at the cost of increased complexity in the estimation of the model parameters. It is found that the perturbed signal model performs better than the harmonic signal model at input signal-to-noise ratios above approximately -10 dB, and that they are equally good below.