LING124 Feature extraction

1
LING124 Feature extraction

• October 14, 2008

2
Class outline

• Feature extraction (Front-end)
• Pre-emphasis
• Framing and windowing
• Mel-Frequency Cepstral Coefficients
• MFSC, MFCC
• Deltas, double-deltas
• Typical feature parameters

3
Feature extraction

• Capture essential acoustic information for speech
  recognition

4
Pre-emphasis

• Increase the magnitude of some range of
  frequencies with respect to other frequencies
• Voiced part of the speech signal is often in high
  frequency range, but high frequency formants have
  smaller amplitude than low frequency formants
• So we apply high-pass filter to speech signal
  before we extract any acoustic features

5
Pre-emphasis (2)

• FIR high-pass filter
• yn xn-axn-1
• 0.9 a 1.0

6
Framing Windowing

• Divide speech signal into successive overlapping
  frames
• Frame-size 10 25 milliseconds
• Frame-shift 5 10 milliseconds
• Multiply individual frames by a specific window
  function
• Hanning window (Nnumber of samples in frame)

7
Spectrum

• Spectrum represents frequency components of
  speech signal
• Spectrum is derived by applying (discrete)
  Fourier transform to a windowed segment of speech
  signal

8
Mel-Frequency Spectral Coefficients

• Perception of pitch is semi-logarithmic
• Recall mel scale and bark scale
• Frequency-dependent smearing
• Ignore distinction that depends on difference
  between two frequencies that are less than 90-120
  mels apart
• Break the spectrum into bands of equal mel
  interval (90 mels)
• Weighted average of spectral coefficients in each
  band
• The log of resulting band averages comprise the
  Mel-Frequency Spectral Coefficients (MFSC)

9
Mel filter bank

• Each filters center frequency follows the
  mel-scale
• mel(frequency) 2595 log10(1frequency/700)
• The edges of a filter coincide with the center
  frequencies of adjacent filters
• maxFreq has to be lower than the Nyquist frequency

10
Mel filter-bank processing
11
Mel-Frequency Cepstral Coefficients

• Discrete Cosine Transform of MFSC
• Represents a signal as a sum of cosine functions
• cf. DFT uses both sine and cosine functions
• This can be interpreted as applying inverse
  discrete Fourier transform to MFSC

12
Delta, double-delta

• Mel-frequency cepstral coefficients do not
  include temporal information
• How spectral shape changes over time is useful in
  speech recognition
• As we normally calculate velocity and
  acceleration, we use first-derivatives (deltas)
  and second-derivatives (double-deltas) of MFCCs
  for temporal information

13
Typical feature parameters

• Pre-emphasis coefficient 0.97
• Frame-size 25ms
• Frame-shift 10ms
• 12 MFCC and log-energy
• Deltas, double-deltas
• A feature vector of 39 components