Week 5

Question 1

What’s the time complexity of the algorithm used for local decoding?

Answer 1

Polynomial time

Question 2

What are the three steps in the local decoding algorithm?

Answer 2

<ul>
<li>Calculate the forward probability</li>
<li>Calculate the backward probability</li>
<li>Combine both to find retrospective distribution</li>
</ul>

Question 3

How do we intuitively decode t?

Answer 3

Fuse information from the past observations and the future observations

Question 4

What does the forward probability calculate?

Answer 4

The sum of probabilities of all paths from given sequence of observable symbols

Question 5

What is the difference between local decoding forward and viterbi algorithm?

Answer 5

Viterbi takes the max where forward algorithm sums

Question 6

What is unsupervised learning?

Answer 6

Estimating the HMMs parameters without annotated tags

Question 7

What is replaced in the Baum Welch algorithm?

Answer 7

The counts are replaced with their estimates

Question 8

What are the three stages in the Baum Welch algorithm?

Answer 8

<ul>
<li>Initialisation</li>
<li>E step</li>
<li>M step</li>
</ul>

Question 9

Describe the initialisation step in the Baum Welch algorithm

Answer 9

Randomly guess some starting value for A0 and B0

Question 10

Describe the E step in the Baum Welch algorithm

Answer 10

<ul>
<li>Apply forward and backward probabilities to calculate</li>
<li>Calculate retrospective distribution</li>
<li>Calculate pseudo transition estimate for all i, q, q'</li>
</ul>

Question 11

Describe the M step in the Baum Welch algorithm

Answer 11

<ul>
<li>Re-estimate A for all q, q'</li>
<li>Re-estimate b for all q,w</li>
<li>Update iteration counter</li>
</ul>

Question 12

How does the likelihood change with each re-estimation step in the Baum Welch algorithm?

Answer 12

It increases, till the likelihood converges for the E and the M step

Question 13

What are the four probabilities for each pair of characters in P(O|C)?

Answer 13

<ul>
<li>rev(i,e)</li>
<li>insp(p,h)</li>
<li>del(k,n)</li>
<li>sub(f,v)</li>
</ul>

Question 14

What does rev(i,e) indicate in misspelling probabilities?

Answer 14

Order has been reversed : ie -> ei

Question 15

What does ins(p,h) indicate in misspelling probabilities?

Answer 15

Second h is inserted after the first letter :

p -> ph

Question 16

What does del(k,n) indicate in misspelling probabilities?

Answer 16

The first letter, preceeding the second is deleted : kn -> n

Question 17

What does sub(f,v) indicate in misspelling probabilities?

Answer 17

The first letter is substituted for the seconds : f -> v

Question 18

What is the spelling correction equation?

Answer 18

argmax P(O|C)P(C)

Question 19

How can we get observable symbols from acoustic waveform?

Answer 19

<ul>
<li>Slice up signal (make it discritised)</li>
<li>Represent each slice as a feature vector</li>
</ul>

Question 20

What does a feature vector for a slice of acoustic noise consist of?

Answer 20

Floating point values representing energy (volume) or frequencies within that slice

Question 21

What is our goal for speech recognition?

Answer 21

Compute the most probable sentence W for the given acoustic observation

Question 22

What are the two main components for automatic speech recognition (ASR)?

Answer 22

<ul>
<li>Language model P(W)</li>
<li>Acoustic model P(O|W)</li>
</ul>

Question 23

What levels may a HMM have when involved with speech instead of written language?

Answer 23

<ul>
<li>Bigram model of words in a sentence</li>
<li>Bigram model of phones within a word</li>
<li>Bigram model of suphones within a phone</li>
</ul>

Question 24

What is a phone?

Answer 24

LANGUAGE INDEPENDENT Individual speech unit

Question 25

What are phones represented by?

Answer 25

Symbols from the phonetic alphabet

Question 26

What does IPA try to do?

Answer 26

Offer sound symbols to transcribe any spoken language with transcription principles

Question 27

What is a phoneme?

Answer 27

Abstract class of sounds that are perceived as one distinctive sound in a given language

Question 28

What is an allophone?

Answer 28

Different pronounciations of the same phoneme

Question 29

What are pronounciation dictionaries?

Answer 29

Tools used for speech recognition and speech synthesis that contain phonemes and their pronounciations

Question 30

How many subphones does a phone model normally distinguish between in ASR?

Answer 30

Three

Question 31

What are the three subphones a phone model distinguishes between in ASR?

Answer 31

Beginning, middle, end

Question 32

How is the variable duration of a subphone represented in a HMM?

Answer 32

Single loop that transitions from the state representing a subphone back to itself

Question 33

What is a Bakis network?

Answer 33

A structure where transitions can only go forwards or loop on a single state, but never backwards

Question 34

How are unresolved ambiguities in speech recognition returned?

Answer 34

Word lattice

Question 35

What is a word lattice

Answer 35

Labelled, directed acyclic graph, with states roughly corresponding to points in time

Question 36

What is the goal of automatic speech recognition?

Answer 36

Computationally build systems that map from an acoustic signal to a string of words

Question 37

What is the intuition of a noisy channel model?

Answer 37

Treat the acuostic waveform as a noisy version of the string of words

Question 38

Describe how the noisy channel model works

Answer 38

If we know how the channel distorts the sound, we could find the correct source sentence for a waveform by taking every possible sentence in a language, running it through channel and seeing if it matches the output

Question 39

What is a HMM characterised by?

Answer 39

<ul>
<li>Set of states</li>
<li>Transition probability matrix, probability of moving from state i to state j</li>
<li>Set of observations</li>
<li>Emission probability matrix, probability of observation being generated from given state</li>
<li>Start and end state</li>
</ul>

Question 40

What are HMM characterized by in speech recognition?

Answer 40

<ul>
<li>Set of states corresponding to subphones</li>
<li>Transition probabilitiy matrix with representations for self loop and going to next subphone</li>
<li>Emission probabilities, probability of a feature vector generated for each subphone state</li>
</ul>

Question 41

What are lexicons?

Answer 41

Lists of words, with a pronounciation for each word expressed as a phone sequence

Question 42

What is the decoding question for speech recognition?

Answer 42

Given a string of acoustic observations, how should we choose the string of words that has the highest posterior probability?