Preemphasis Influence on Harmonic Speech Model with Autoregressive Parameterization

Abstract

Autoregressive speech parameterization with and without preemphasis is discussed for the source-filter model and the harmonic model. Quality of synthetic speech is compared for the harmonic speech model using autoregressive parameterization without preemphasis, with constant and adaptive preemphasis. Experimental results are evaluated by the RMS log spectral measure between the smoothed spectra of original and synthesized male, female, and childish speech sampled at 8 kHz and 16 kHz. Although the harmonic model is used, the benefit of the adaptive preemphasis could be valid for the source-filter model, as well.