SP Module 8 Speech Recognition & Feature Engineering

Gaussian distributions in models

Classification model

Gaussian distribution of classification result of feature vector

Decision boundary

Cepstral Analysis, Mel-Filterbanks

We now start thinking about what a good representation of the acoustic signal should be, motivating the use Mel-Frequency Cepstral Coefficients (MFCCs).

Since the feature in a feature vector is correlated, if we want to use gaussian, we have to solve this correlation problem.

In TTS

In ASR, we can decompose the input further

The independent variable of a cepstral graph is called the quefrency. The quefrency is a measure of time, though not in the sense of a signal in the time domain. Then the first 12 value in cepstrum is our result feature vector, after feature engineering.

Then we can solve the correlation problem discussed in previous, by using MFCCs.

MFCCs

Overview of steps required to derive MFCCs, moving towards modelling MFCCs with Gaussians and Hidden Markov Models