Example: LPC10

So called because 10 LP coefficients are used.

• 8 kHz sample rate, 180 samples/frame, 44.44 frames/second
• Order 10 LP analysis: First two coefficients are quantised as log area ratios with five bits each, last 8 as reflection coefficients. Number of bits per coefficient decreases with index down to two bits
• 7 bits used for pitch and voicing decision
• 5 bits used for gain
• Total: 54 bits per frame, 2400 bps

Disadvantages:

• Can sound very ``buzzy'' - noise through parameter updates
• Poor LP modelling results in wide bandwidths and rapid decay of the pulse excitation
• Regularly voiced excitation is unnatural - normally some jitter
• Voicing errors produce significant distortions
• Binary voicing decision is sometimes poor
• Not suited to model nasals - although okay in practice
• Only models speech - doesn't work if background noise (i.e. not suited to mobile phone applications without further work).

Items which have to be addressed:

• Glottal pulse shaping
• Pitch synchronous parameter updating
• Fine tuning of the voicing decision
• Separation of speech and noise
• Exploitation of temporal correlations of acoustic vectors