Audio Signal Processing and Coding

4.4 OPEN-LOOP ANALYSIS-SYNTHESIS LINEAR PREDICTION

In almost all LP-based speech codecs, speech is approximated on short analysis intervals, typically in the neighborhood of 20 ms. As shown in Figure 4.6, a set of LP synthesis parameters is estimated on each analysis frame to capture the shape of the vocal tract envelope and to model the excitation.

Some of the typical synthesis parameters encoded and transmitted in the open-loop LP include the prediction coefficients, the pitch information, the frame energy, and the voicing. At the receiver, the transmitted “source” parameters are used to form the excitation. The excitation, e(n), is then used to excite the LP synthesis filter, 1/A(z), to reconstruct the speech signal. Some of the standardized open-loop analysis-synthesis LP algorithms include the LPC10e Federal Standard FS-1015 [FS1015] [Trem82] [Camp86] and the Mixed Excitation LP (MELP) [McCr91]. The LPC10e FS-1015 uses a tenth-order predictor to estimate the vocal tract parameters and a two-state voiced or unvoiced excitation model for residual modeling. Mixed excitation schemes in conjunction with LPC were proposed by Makhoul et al. [Makh78] and were later revisited by McCree and Barnwell [McCr91] [McCr93].

Figure 4.5. The ADPCM ITU-T G.726 encoder.

Figure 4.6. Open-loop analysis-synthesis ...