Noise suppression properties of synthesis error optimization

Analysis-by-synthesis (AbS) speech codecs such as multi-pulse LPC and CELP compute the coefficients of the synthesis filter by minimizing the linear prediction (LP) error and the excitation function by minimizing the closed-loop synthesis error. In the presence of additive noise, the LPC coefficients will be affected by noise, resulting in a synthesized speech with degraded quality. This paper points out the noise advantages of computing both the filter parameters and the excitation function by minimizing the closed-loop synthesis error. By minimizing the synthesis error, the analysis and synthesis stages become more compatible. Furthermore, It is shown that unlike LP-based estimation, minimization of the synthesis error could lead to noise-free filter parameters for uncorrelated (colored) noise. An approach for reducing the synthesis error is presented and simulation results on both synthetic and real speech are given.