You are here:Home Courses Archived courses Speech Processing (up to 2016-17) Automatic speech recognition Mel frequency cepstral coefficients

Mel frequency cepstral coefficients

Another common step in machine learning is to use our knowledge to engineer a better parameterisation of the signal.

  • Requirements

    We've decided to use Gaussian pdfs, so that places some requirements on the features that we model.

  • Feature engineering

    Those requirements can be imposed by a little clever engineering.

  • Filterbank

    The filterbank is the first step in feature engineering: it warps the frequency scale and removes F0.

  • The complete process

    A sketch of the complete process.

Reading

Holmes & Holmes – Chapter 10 – Front-end analysis for ASR

Covers filterbank, MFCC features. The material on linear prediction is out of scope.

Taylor – Section 12.3 – The cepstrum

By using the logarithm to convert a multiplication into a sum, the cepstrum separates the source and filter components of speech.

Ladefoged (Elements) – Chapter 6 – Hearing

Some understanding of human hearing will be helpful for engineering suitable features to extract from the waveform for automatic speech recognition.