Introduction to Speech Recognition

Question 1

What probability model will we construct? How does this work

Answer 1

Hidden Markov Model.

Construct a statistical model of how signals are generated and use probability calculus in order to update of belief.

Question 2

What is a sound wave? How do we extract useful features?

Answer 2

A measure of a change in air pressure over time. Signal processing.

1. break signal up into a sequence of overlapping segments
2. Use a Fourier transform to extract the dominant frequencies of the signal.
3. Obtain a set of Mel-frequency cepstrum coefficients

Question 3

What is an MFCC?

Answer 3

Mel-frequency cepstrum coefficients.
Numbers representing the contribution from different frequency bands obtained by Fourier transform.
Each segmented part of the speech signal has a vector of (13) MFCC features

Question 4

What is a phoneme? What do they help us to do?

Answer 4

small elementary utterances. Help us to represent longer words

Question 5

What is a statistical language model?

Answer 5

Disambiguate cases by identifying simple patterns in language. i.e. which word is likely to follow another.

Question 6

How do we combine language models with evidence fro speech signals?

Answer 6

Probabilistic methods

Question 7

What is training?

Answer 7

process whereby the parameters of a model are adapted to a particular problem domain

Question 8

How can sensory information be represented in useful ways?

Answer 8

Sound waves

Robust sensor data