Analysis by synthesis speech coding with generalized pitch prediction
Paul G. Mermelstein, Yasheng Qian · 1999
A new analysis-by-synthesis speech coding structure is presented for high-quality speech coding in the 4 to 8 kb/s range. CELP with generalized pitch prediction (GPP-CELP) differs from classical code-excited linear prediction (CELP) in that for voiced segments it is the speech signal that is decomposed into a component predictable with the aid of the adaptive codebook (ACB) and a nonpredictable aperiodic component, not the LPC residual. The spectrum of the aperiodic component is estimated by linear-prediction analysis. An approximation to the aperiodic component is synthesized from a stochastic codebook of sparse pulse sequences and its spectrum is shaped by the LPC synthesis filter. The ACB contains samples of the past reconstructed signal, low-passed to increase the pitch prediction gain. For voiced segments the new structure yields higher pitch prediction gain and lower linear-prediction gain than classical CELP. Subjective and objective comparisons reveal significant advantages for GPP-CELP over classical CELP.