Lecture 6 - Feature Extraction

Question 1

Give two applications of feature extraction?

Answer 1

1. Classification; leave only relevant information
   - speech recognition, speaker recognition,
2. Coding; leave only perceptually (who is speaking) important information
   - Real-time transmission of speech
   - Online content

Question 2

The cepstral deconvolution separates the speech signal into components varying slowly (envelope) and more rapidly (excitation). Which coefficients do we keep?

Answer 2

Usually keep the lowest 10-15 coefficients; related to vocal tract resonances and keep the ‘filter’.

Question 3

Speech has more energy at lower frequencies and therefore contain numerical problems in implementation. What can be done in order to boost higher frequencies?

Answer 3

Use a time-domain FIR filter with one free parameter
- P(z) = 1 - az^-1, a depends on the sampling frequency.

Can make information in higher formants more available to the acoustic model