Hidden Markov Models I

Recognition of isolated words

Recommended texts:
Spoken Language Processing, Chapter 8
Speech and Language Processing, Appendix A
 Statistical recognition
• I. Recognition of isolated words
– Training: creation of a model for each word.
– Recognition: To determine the best match
model/utterance

• II. Large Vocabulary Speech Recognition (LVSR)

– Subword units (phones) in context, e.g., m-a+r
– Phonetic dictionary
– Language model P(house | the white)
 Isolated words. General scheme

Model W1

Model W2 Training

speech
Model WN
Parameterizer recognized
speech

Classifier Decision

Recognition
 Hidden Markov Model (HMM)
T11 T22 T33

T12 T23

T13

• Generic representation of a statistical model for processes that

generate time series. The HMM is a sequence model.
• The “segments” in the time series are referred to as states: the
process passes through these states to generate time series
• The entire structure may be viewed as a generalization of DTW
models
Bhiksha Raj and Rita Singh (CMU)
 Hidden Markov Models

• A Hidden Markov Model consists of two components

– A state/transition backbone that specifies how many states there
are, and how they can follow one another
– A set of probability distributions, one for each state, which
specifies the distribution of all vectors in that state

Markov chain

Data distributions

HMMs
 HMM as a statistical model
• An HMM is a statistical model for a time-varying process
• The process is always in one of a countable number of states at
any time

• When the process visits in any state, it generates an

observation by a random draw from a distribution associated
with that state

• The process constantly moves from state to state. The

probability that the process will move to any state is
determined solely by the current state
– i.e. the dynamics of the process are Markovian

• The entire model represents a probability distribution over the

sequence of observations
– It has a specific probability of generating any particular sequence
HMMs
– The probabilities of all possible observation sequences sums to 1
 How an HMM models a process

HMM assumed to be
generating data

state sequence

state distributions

Observation
sequence

HMMs
 HMMs are abstractions
• The states are not directly observed
– Here states of the process are analogous to configurations of the vocal tract that
produces the signal
– We only hear the speech; we do not see the vocal tract
– i.e. the states are hidden

• The interpretation of states is not always obvious

– The vocal tract actually goes through a continuum of configurations
– The model represents all of these using only a fixed number of states

• The model abstracts the process that generates the data

– The system goes through a finite number of states
– When in any state it can either remain at that state, or go to another with some
probability
– When at any states it generates observations according to a distribution associated with
that state

17 March 2007 HMMs

 HMM Parameters
0.6 0.7
• The topology of the HMM
0.4

– No. of states and allowed

transitions 0.5 0.3

– E.g. here we have 3 states and

cannot go from the blue state to
the red 0.5

• The transition probabilities æ .6 .4 0 ö

ç ÷
– Often represented as a matrix as T = ç 0 .7 .3 ÷
here ç .5 0 .5 ÷
è ø
– Tij is the probability that when in
state i, the process will move to j
• The probability of beginning at a
particular state
• The state output distributions

17 March 2007 HMMs

 HMM state output distributions
• The state output distribution represents the distribution of data produced from
any state. We can have:
• Discrete probabilities (DHMM) , e.g., Vector Quantization (k-means)
• Continuous probabilities , e.g., Gaussian Mixture Model (GMM), DNN

Bhiksha Raj and Rita Singh (CMU)

 Discrete Hidden Markov models
a 22
a11 a33
a13
a12 b3(ot)
b1(ot) a23
1 a21
2 a32
3
• N states {1...N} b2(ot) a31
• At instant of time t the system will be in a specific state si

• In equispaced intervals, the system may change its state with probability aij
aij = P(qt = j | qt-1= i) aij>0,

• Every time a transition occurs, the system generates an observation from a

finite alphabet that depends on the state to which has moved.
b j (Ot ) = P (Ot qt = j )
• There is no longer a one-to-one correspondence between the observation
sequence and the state sequence, so you cannot unanimously determine the
state sequence for a given observation sequence; i.e., the state sequence is
not observable and therefore hidden.
 Types of models
a
11

Ergodic 1
Phonotactic recognition
a12
a
13

a a31
21

a
23

a22 2 3 a 33

a32

Left-right or Backis’ Phonetic units modeling

a a a a a
11 22 33 44 nn

a a a a
1 12 2 23 3 34 4 (n-1)n n
 Problems to be solved
• Evaluation: Given a sequence of observations O = o1
o2...oT and a model l, what is the probability P(O|l)
that the model generates the observations?.
• Decoding: Given a sequence of observations O = o1
o2...oT and a model l, find the optimum sequence of
states Q=q1,...,qT
• Training: Given the model l=(p,A,B) and a set of
training sequences, how to adjust the model
parameters l to maximize the joint probability P(O|l)
 Evaluation
Given the sequence O = o1 o2...oT and the model
l= (p,A,B), calculate P(O)
Solution: Let’s suppose a sequence of states:
Q= q1q2...qT, that have created the observations.

P(Q) = π q aq q ...aq
1 1 2 T −1qT

P(O,Q) = π q bq (o1 )aq q bq (o2 )...aq bq (oT )

1 1 1 2 2 T −1qT T

P(O) = ∑ π q bq (o1 )aq q bq (o2 )...aq b (oT )

1 1 1 2 2 T −1qT qT
all Q
 Efficient solution
• The direct method requires 2TNT operations
1sec:100 observations and 5 states: 1072 calculations

• Forward backward algorithm

– Accumulates operations in sequences with repeated paths
– #operations N2T = 2500 operations
 Most likely sequence of states

• Given a sequence of observations, searches, in

every instant t, the most likely sequence of states

• Viterbi’s algorithm

• #operations: N2T (multiplications or additions)

 Viterbi’s algorithm
Sequence of states with the highest probability
dt(j) = max P(q1 q2 ...qt-1 qt = j , o1 o2 ...ot )
q1,q2...qt-1

Initialization d1(i) = pi bi(o1)

y1(i) = 0 1£i£N

Recursion: dt(j) = max [dt-1(i) aij] bj(ot) 2£t£T

1£i£N

yt(j) = arg max [dt-1(i) aij] 1£j£N

1£i£N

Ending P* = max [dT(i)]

1£i£N

qT*=arg max [dT(i)]

1£i£N

Sequence (backtracking)
qt* = yt+1( qt+1*) t = T-1, T-2, ...1
State Initialization

4
3
2
1
1 2 T signal
Initialization: For every
state, we calculate the
pjbj(O1) probability of generating
the first observation.
State Recursion
4
3
2
1
1 2 T signal
dt-1(4) a42 In t, we calculate the highest
a32 probability of arriving at the state
n from the others and generating
a22 dt(2) the observation Ot
dt-1(1) a12 In yt(n) we note down the state
from which we achieve the
highest probability
b2(Ot)
 Ending: backtracking

State

4
3
2
1
1 2 T signal

Sequence (backtracking)
qt* = yt+1( qt+1*) t = T-1, T-2, ...1
 Training or parameter
estimation

• Given the sequence O = o1 o2... oT adjust the model

l= (p,A,B) that maximizes P(O|l)
• Iterative solution: Baum Welch
– start from an initial model
– pass the training sequence and re-estimate the model
parameters
– P(O|l) > P(O| l)
 Initialization
• Manual
• Automatic
– Uniform matrix A
– Segmentation of the training sequences into number of
states
– Grouping of the spectra in every state
– Estimation of B
– Iterate until convergence
• Iterate: using Viterbi algorithm to segment the training sequence into states
improves the initialization
 Isolated words. General scheme

HMM λ0 P(O|λ0)
speech

O = o1 o2...oT HMM λ1 P(O|λ1)

Parameterizer

HMM λM P(O|λM)

• Each HMM model is trained with its own set of recordings of isolated words.
• Each test recording has a single isolated word
• P(O|λi) can be computed by the forward algorithm. In practice, the Viterbi
algorithm ( max P(Q,O|λi) ) is faster and gives the same accuracy.