ECE 259 Digital Speech Processing LPC vocoder project

1
ECE 259Digital Speech ProcessingLPC vocoder
project

• Teacher Prof Lawrence Rabiner
• Student Vinay Melkote

2
LPC analysis

• Predict speech sample s(n) from p previous
  samples with error minimisation.
• s(n) S ai s(n-i) e(n) i1,2,,p
• Exploits correlation in speech

3
Philosophy

• Vocal tract can be modelled as loseless tubes
• The transfer function has poles and zeros
• H(z) N(z)/D(z)
• The coefficients of the denominator D(z)
  polynomial are ai s and determine the resonances
  or formants

4
Finding ai s

• Since the ai characterise formants, they vary
  with time as the utterances change
• Assumed constant over a frame of speech
• To find ai minimise error over that frame
• Windowed speech - autocorrelation method
• Windowed error covariance method

5
Lattice implementation

• On the fly method
• Find reflection coefficients
• Use Durbins recursion to find ai

6
LPC synthesis

• Based on the following model of speech production

Figure from Digital Processing of Speech Signals
, by Lawrence Rabiner and Ronald Schafer
7
LPC Vocoder

• Analyse get ai s and gains
• Synthesize reconstruct error signal, extract
  window and filter through vocal tract filter,
  overlap add

8
The Matlab code
waveanalx
vinlpcx
frameconst
Pitch contour
u(n)
Gain compute
Error sum
synthesizex
Reconstructed speech
9
Synthesized speech samples

• Original
• N240, L80,p 12
• N generated
• L generated
• Hamming window

10
Gain quantized

• Original
• No of bits
• 16
• 12
• 8
• 4
• 2
• 1

11
Log area parameter quantized

• Alpha quantized
• Log area parameter quantized
• No of bits
• 12
• 8
• 6
• 4
• 2

12
Log area and gain quantized

• No of bits No of bits
• for log area for gain
• 8 8
• 8 4
• 8 2
• 6 8
• 6 4
• 6 2
• 4 8
• 4 4
• 4 2

13
Conclusion

• LPC synthesis gives speech which is highly
  intelligible and of good quality
• Quantizing the required gain and vocal tract
  parameters can still maintain intelligibility and
  appreciable quality
• Thus LPC helps to achieve low bit rate coding of
  speech