Parametrisation: Sophistication

1
Parametrisation Sophistication

• We need something more representative of the
  information in the speech less prone to variation
• The spectral slices we have been viewing to date
  in Praat are actually LPC (Linear Predictive
  Coding) spectra
• LPC attempts to remove the effects of phonation
• Leaves us with correlate of VT configuration

2
Spectral Feature Extraction

• Extract compact set of spectral parameters
  (features) for each frame
• Frames usually overlapping

3
DFT spectra vs LPC spectra

• DFT (Discrete Fourier Transform)
• Technique ubiquitous in DSP for spectral analysis
• fft function in MATLAB
• demo gt Numericsgt Fast Fourier Transform
• LPC
• Mathematical encoding of signals
• Based on modelling speech as a series of sums of
  exponentially decaying sinusoids
• Source-filter decomposition
• Typical example of how spectral information can
  be compressed

4
Estimating Spectra

• Choose frequency resolution
• Time/Frequency trade off
• Parametrisation frame length
• Pre-emphasise
• Flattens spectrum which reduces spectral dynamic
  range which eases estimation
• Apply window function in time domain
• Tapers frame boundary values to zero
• Gives better picture of spectrum

5
DFT Spectrum /u/
6
Frame Length5,40,200ms
7
Freq. Resolution for 5,40,200ms
8
Preemphasis using diff
9
Preemphasis
10
Windowing using hamming
11
Spectral Effect
12
LPC Spectrum using lpc
13
LPC

• Linear Predictive Coding
• Rule of thumb for order
• (kHz of Sampling Frequency) (2 to 4)
• In previous figure, order 14 was used
• LP Coefficients can be easily transformed to
  centre frequencies and bandwidths of peaks in
  spectrum
• MATLAB lpc
• 1st coefficient returned always 1, so omit

14
Cepstrally Smoothed Spectrum