Lecture 7 - ML

Question 1

Give applications for Gaussian Mixture Models (GMMs) in speech applications

Answer 1

1. Speaker recognition
2. Emotion classification
3. Voiced/unvoiced decisions

Question 2

Why is a HMM preferred over a GMM

Answer 2

Using a GMM, there is no way to capture time. Time is an important aspect in speech and therefore, use Hidden Markov Models (HMMs)

Question 3

A gaussian or normal distribution is common and easily analysed. To estimate the parameters, the sample mean and sample variance can be used. What’s the disadvantage of this?

Answer 3

This is called the frequentist approach.

-> doesn’t really work for large and complex training data.

Question 4

It can be assumed that the matrix has diagonal covariance. What does this mean?

Answer 4

This means that the features are independent.