Model based speech coding consists in modeling the vocal tract as a linear time-variant system. The all-pole model produced by the computationally efficiency Linear Predictive Coding method provides a good representation for the majority of speech sounds. However, nasal and fricative sounds, as well as stop consonants, contain spectral zeros which requires the use of a zero-pole model. Roughly speaking, a zero-pole model estimation method typically does a non-parametric estimation of the vocal tract impulse response, and tunes the zero-pole model to fit this estimation in a square sense. In this paper we propose an alternative strategy. We tune the zero-pole model to directly fit the power spectrum of the speech signal in a logarithmic scale, to be consistent with the way the human ear perceives sounds. In this way, we avoid the error introduced by the vocal tract impulse response estimation and obtain a model that is more accurate at reproducing spectral zeros in a logarithmic scale. A drawback of the proposed method, however, is its computational complexity.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.