Linear prediction vocoders

• Employ the source filter model (e.g. figure 9).
• For ever frame, need to code:
  • A representation of the LP filter
  • Power
  • Degree of voicing
  • Pitch (if voiced)
• Most bits go into the LP parameters

Commonly used representations of the LP parameters:

• LP coefficients: when quantisation is not a problem
• Reflection coefficients: robust but not very efficient
• Line spectral pairs: most efficient

Quantisation:

• Independent: about 50 bits per frame
• Vector quantisation: about 25 bits per frame

  
Figure 46: Simple LPC vocoder

 
• Parameter updating
• Example: LPC10