Hidden Markov Models and Gaussian Mixture Models

Size: px
Start display at page:

Download "Hidden Markov Models and Gaussian Mixture Models"

Transcription

1 Hidden Markov Models and Gaussian Mixture Models Hiroshi Shimodaira and Steve Renals Automatic Speech Recognition ASR Lectures 4&5 23&27 January 2014 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 1

2 Overview HMMs and GMMs Key models and algorithms for HMM acoustic models Gaussians GMMs: Gaussian mixture models HMMs: Hidden Markov models HMM algorithms Likelihood computation (forward algorithm) Most probable state sequence (Viterbi algorithm) Estimting the parameters (EM algorithm) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 2

3 Fundamental Equation of Statistical Speech Recognition If X is the sequence of acoustic feature vectors (observations) and W denotes a word sequence, the most likely word sequence W is given by W = arg max P(W X) W ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 3

4 Fundamental Equation of Statistical Speech Recognition If X is the sequence of acoustic feature vectors (observations) and W denotes a word sequence, the most likely word sequence W is given by Applying Bayes Theorem: W = arg max P(W X) W P(W X) = p(x W)P(W) p(x) p(x W)P(W) W = arg max W p(x W) }{{} Acoustic model P(W) }{{} Language model ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 3

5 Acoustic Modelling Recorded Speech Signal Analysis Decoded Text (Transcription) Hidden Markov Model Acoustic Model Training Data Lexicon Language Model Search Space ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 4

6 Hierarchical modelling of speech "No right" Utterance NO RIGHT Word n oh r ai t Subword HMM Acoustics ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 5

7 Hierarchical modelling of speech Generative Model "No right" Utterance NO RIGHT Word n oh r ai t Subword HMM Acoustics ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 5

8 Acoustic Model: Continuous Density HMM P(s 1 s 1 ) P(s 2 s 2 ) P(s 3 s 3 ) s 1 s 2 s 3 s E s I P(s 1 s I ) P(s 2 s 1 ) P(s 3 s 2 ) P(s E s 3 ) p(x s 1 ) p(x s 2 ) p(x s 3 ) x x x Probabilistic finite state automaton Paramaters λ: Transition probabilities: a kj = P(S = j S = k ) Output probability density function: b j (x) = p(x S = j ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 6

9 Acoustic Model: Continuous Density HMM s I s 1 s 2 s 3 s E Paramaters λ: x 1 x 2 x 3 x 4 x 5 x 6 Probabilistic finite state automaton Transition probabilities: a kj = P(S = j S = k ) Output probability density function: b j (x) = p(x S = j ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 6

10 HMM Assumptions P(s 1 s 1 ) P(s 2 s 2 ) P(s 3 s 3 ) s 1 s 2 s 3 s E s I P(s 1 s I ) P(s 2 s 1 ) P(s 3 s 2 ) P(s E s 3 ) p(x s 1 ) p(x s 2 ) p(x s 3 ) x x 1 Observation independence An acoustic observation x is conditionally independent of all other observations given the state that generated it 2 Markov process A state is conditionally independent of all other states given the previous state x ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 7

11 HMM Assumptions s(t 1) s(t) s(t+1) x(t 1) x(t) x(t + 1) 1 Observation independence An acoustic observation x is conditionally independent of all other observations given the state that generated it 2 Markov process A state is conditionally independent of all other states given the previous state ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 8

12 HMM OUTPUT DISTRIBUTION ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 9

13 Output distribution P(s 1 s 1 ) P(s 2 s 2 ) P(s 3 s 3 ) s 1 s 2 s 3 s E s I P(s 1 s I ) P(s 2 s 1 ) P(s 3 s 2 ) P(s E s 3 ) p(x s 1 ) p(x s 2 ) p(x s 3 ) x x x ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 10

14 Output distribution P(s 1 s 1 ) P(s 2 s 2 ) P(s 3 s 3 ) s 1 s 2 s 3 s E s I P(s 1 s I ) P(s 2 s 1 ) P(s 3 s 2 ) P(s E s 3 ) p(x s 1 ) p(x s 2 ) p(x s 3 ) x x x Single multivariate Gaussian with mean µ j, covariance matrix Σ j : b j (x) = p(x j ) = N (x; µ j, Σ j ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 10

15 Output distribution P(s 1 s 1 ) P(s 2 s 2 ) P(s 3 s 3 ) s 1 s 2 s 3 s E s I P(s 1 s I ) P(s 2 s 1 ) P(s 3 s 2 ) P(s E s 3 ) p(x s 1 ) p(x s 2 ) p(x s 3 ) x x x Single multivariate Gaussian with mean µ j, covariance matrix Σ j : b j (x) = p(x j ) = N (x; µ j, Σ j ) M-component Gaussian mixture model: M b j (x) = p(x j ) = c jm N (x; µ jm, Σ jm ) m=1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 10

16 Output distribution P(s 1 s 1 ) P(s 2 s 2 ) P(s 3 s 3 ) s 1 s 2 s 3 s E s I P(s 1 s I ) P(s 2 s 1 ) P(s 3 s 2 ) P(s E s 3 ) p(x s 1 ) p(x s 2 ) p(x s 3 ) x x x Single multivariate Gaussian with mean µ j, covariance matrix Σ j : b j (x) = p(x j ) = N (x; µ j, Σ j ) M-component Gaussian mixture model: M b j (x) = p(x j ) = c jm N (x; µ jm, Σ jm ) m=1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 10

17 Background: cdf Consider a real valued random variable X Cumulative distribution function (cdf) F (x) for X : F (x) = P(X x) To obtain the probability of falling in an interval we can do the following: P(a < X b) = P(X b) P(X a) = F (b) F (a) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 11

18 Background: pdf The rate of change of the cdf gives us the probability density function (pdf), p(x): p(x) = d dx F (x) = F (x) F (x) = x p(x)dx p(x) is not the probability that X has value x. But the pdf is proportional to the probability that X lies in a small interval centred on x. Notation: p for pdf, P for probability ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 12

19 The Gaussian distribution (univariate) The Gaussian (or Normal) distribution is the most common (and easily analysed) continuous distribution It is also a reasonable model in many situations (the famous bell curve ) If a (scalar) variable has a Gaussian distribution, then it has a probability density function with this form: ( ) p(x µ, σ 2 ) = N(x; µ, σ 2 1 (x µ) 2 ) = exp 2πσ 2 2σ 2 The Gaussian is described by two parameters: the mean µ (location) the variance σ 2 (dispersion) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 13

20 Plot of Gaussian distribution Gaussians have the same shape, with the location controlled by the mean, and the spread controlled by the variance One-dimensional Gaussian with zero mean and unit variance (µ = 0, σ 2 = 1): mean=0 variance=1 pdf of Gaussian Distribution p(x m,s) x ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 14

21 Properties of the Gaussian distribution N(x; µ, σ 2 ) = ( ) 1 (x µ) 2 exp 2πσ 2 2σ pdfs of Gaussian distributions 0.35 mean=0 variance=1 0.3 mean=0 variance= p(x m,s) mean=0 variance= x ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 15

22 Parameter estimation Estimate mean and variance parameters of a Gaussian from data x (1), x (2),..., x (N) Use sample mean and sample variance estimates: µ = 1 N σ 2 = 1 N N n=1 x (n) (sample mean) N (x (n) µ) 2 (sample variance) n=1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 16

23 Exercise maximum likelihood estimation (MLE) Consider the log likelihood of a set of N training data points {x (1),..., x (N) } being generated by a Gaussian with mean µ and variance σ 2 : L = ln p({x (1),..., x (n) } µ, σ 2 ) = 1 2 = 1 2σ 2 N n=1 ( (x (n) µ) 2 σ 2 ln σ 2 ln(2π) N (x (n) µ) 2 N 2 ln σ2 N 2 ln(2π) n=1 By maximising the the log likelihood function with respect to µ show that the maximum likelihood estimate for the mean is indeed the sample mean: µ ML = 1 N x (n). N n=1 ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 17

24 The multivariate Gaussian distribution The d-dimensional vector x = (x 1,..., x d ) T follows a multivariate Gaussian (or normal) distribution if it has a probability density function of the following form: ( 1 p(x µ, Σ) = (2π) d/2 exp 1 ) Σ 1/2 2 (x µ)t Σ 1 (x µ) The pdf is parameterized by the mean vector µ = (µ 1,..., µ d ) T σ σ 1d and the covariance matrix Σ = σ d1... σ dd ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 18

25 The multivariate Gaussian distribution The d-dimensional vector x = (x 1,..., x d ) T follows a multivariate Gaussian (or normal) distribution if it has a probability density function of the following form: ( 1 p(x µ, Σ) = (2π) d/2 exp 1 ) Σ 1/2 2 (x µ)t Σ 1 (x µ) The pdf is parameterized by the mean vector µ = (µ 1,..., µ d ) T σ σ 1d and the covariance matrix Σ = σ d1... σ dd The 1-dimensional Gaussian is a special case of this pdf ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 18

26 The multivariate Gaussian distribution The d-dimensional vector x = (x 1,..., x d ) T follows a multivariate Gaussian (or normal) distribution if it has a probability density function of the following form: ( 1 p(x µ, Σ) = (2π) d/2 exp 1 ) Σ 1/2 2 (x µ)t Σ 1 (x µ) The pdf is parameterized by the mean vector µ = (µ 1,..., µ d ) T σ σ 1d and the covariance matrix Σ = σ d1... σ dd The 1-dimensional Gaussian is a special case of this pdf The argument to the exponential 0.5(x µ) T Σ 1 (x µ) is referred to as a quadratic form. ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 18

27 Covariance matrix The mean vector µ is the expectation of x: µ = E[x] ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 19

28 Covariance matrix The mean vector µ is the expectation of x: µ = E[x] The covariance matrix Σ is the expectation of the deviation of x from the mean: Σ = E[(x µ)(x µ) T ] ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 19

29 Covariance matrix The mean vector µ is the expectation of x: µ = E[x] The covariance matrix Σ is the expectation of the deviation of x from the mean: Σ = E[(x µ)(x µ) T ] Σ is a d d symmetric matrix: σ ij = E[(x i µ i )(x j µ j )] = E[(x j µ j )(x i µ i )] = σ ji ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 19

30 Covariance matrix The mean vector µ is the expectation of x: µ = E[x] The covariance matrix Σ is the expectation of the deviation of x from the mean: Σ = E[(x µ)(x µ) T ] Σ is a d d symmetric matrix: σ ij = E[(x i µ i )(x j µ j )] = E[(x j µ j )(x i µ i )] = σ ji The sign of the covariance helps to determine the relationship between two components: ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 19

31 Covariance matrix The mean vector µ is the expectation of x: µ = E[x] The covariance matrix Σ is the expectation of the deviation of x from the mean: Σ = E[(x µ)(x µ) T ] Σ is a d d symmetric matrix: σ ij = E[(x i µ i )(x j µ j )] = E[(x j µ j )(x i µ i )] = σ ji The sign of the covariance helps to determine the relationship between two components: If x j is large when x i is large, then (x i µ i )(x j µ j ) will tend to be positive; ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 19

32 Covariance matrix The mean vector µ is the expectation of x: µ = E[x] The covariance matrix Σ is the expectation of the deviation of x from the mean: Σ = E[(x µ)(x µ) T ] Σ is a d d symmetric matrix: σ ij = E[(x i µ i )(x j µ j )] = E[(x j µ j )(x i µ i )] = σ ji The sign of the covariance helps to determine the relationship between two components: If x j is large when x i is large, then (x i µ i )(x j µ j ) will tend to be positive; If x j is small when x i is large, then (x i µ i )(x j µ j ) will tend to be negative. ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 19

33 Spherical Gaussian Surface plot of p(x 1, x 2 ) 2 Contour plot of p(x 1, x 2 ) p(x 1, x 2 ) x x µ = ( 0 0 x 1 ) Σ = x 1 ( ) ρ 12 = 0 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 20

34 Diagonal Covariance Gaussian Surface plot of p(x 1, x 2 ) 4 Contour plot of p(x 1, x 2 ) p(x 1, x 2 ) x x 2 2 µ = 4 4 ( ) x 1 Σ = x 1 ( ) ρ 12 = 0 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 21

35 Full covariance Gaussian Surface plot of p(x 1, x 2 ) 4 Contour plot of p(x 1, x 2 ) p(x 1, x 2 ) x x 2 µ = 2 4 ( ) 2 1 x Σ = ( x 1 ) ρ 12 = 0.5 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 22

36 Parameter estimation of a multivariate Gaussian distribution It is possible to show that the mean vector ˆµ and covariance matrix ˆΣ that maximize the likelihood of the training data are given by: ˆµ = 1 N ˆΣ = 1 N N n=1 x (n) N (x (n) ˆµ)(x (n) ˆµ) T n=1 The mean of the distribution is estimated by the sample mean and the covariance by the sample covariance ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 23

37 Example data 10 5 X X1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 24

38 Maximum likelihood fit to a Gaussian 10 5 X X1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 25

39 Data in clusters (example 1) µ 1 = [0 0] T µ 2 = [1 1] T Σ 1 = Σ 2 = 0.2 I ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 26

40 Example 1 fit by a Gaussian µ 1 = [0 0] T µ 2 = [1 1] T Σ 1 = Σ 2 = 0.2 I ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 27

41 k-means clustering k-means is an automatic procedure for clustering unlabelled data Requires a prespecified number of clusters Clustering algorithm chooses a set of clusters with the minimum within-cluster variance Guaranteed to converge (eventually) Clustering solution is dependent on the initialisation ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 28

42 k-means example: data set (4,13) 10 (2,9) (7,8) 5 (6,6) (7,6) (4,5) (5,4) (8,4) (10,5) 0 (1,2) (5,2) (1,1) (3,1) (10,0) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 29

43 k-means example: initialization (4,13) 10 (2,9) (7,8) 5 (6,6) (7,6) (4,5) (5,4) (8,4) (10,5) 0 (1,2) (5,2) (1,1) (3,1) (10,0) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 30

44 k-means example: iteration 1 (assign points to clusters) (4,13) 10 (2,9) (7,8) 5 (6,6) (7,6) (4,5) (5,4) (8,4) (10,5) 0 (1,2) (5,2) (1,1) (3,1) (10,0) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 31

45 k-means example: iteration 1 (recompute centres) (4,13) 10 (2,9) (4.33, 10) (7,8) 5 0 (6,6) (7,6) (4,5) (10,5) (5,4) (8,4) (8.75,3.75) (3.57, 3) (1,2) (5,2) (1,1) (3,1) (10,0) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 32

46 k-means example: iteration 2 (assign points to clusters) (4,13) 10 (2,9) (4.33, 10) (7,8) 5 0 (7,6) (6,6) (4,5) (10,5) (5,4) (8,4) (8.75,3.75) (3.57, 3) (1,2) (5,2) (1,1) (3,1) (10,0) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 33

47 k-means example: iteration 2 (recompute centres) (4,13) 10 (2,9) (4.33, 10) (7,8) 5 (6,6) (7,6) (4,5) (5,4) (8,4) (10,5) (8.2,4.2) 0 (1,2) (1,1) (3.17, 2.5) (5,2) (3,1) (10,0) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 34

48 k-means example: iteration 3 (assign points to clusters) (4,13) 10 (2,9) (4.33, 10) (7,8) 5 (4,5) (6,6) (5,4) (7,6) (8,4) (10,5) (8.2,4.2) 0 (1,2) (1,1) No changes, so converged (3.17, 2.5) (5,2) (3,1) (10,0) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 35

49 Mixture model A more flexible form of density estimation is made up of a linear combination of component densities: p(x) = M p(x j )P( j ) j=1 This is called a mixture model or a mixture density p(x j ): component densities P( j ): mixing parameters Generative model: 1 Choose a mixture component based on P( j ) 2 Generate a data point x from the chosen component using p(x j ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 36

50 Component occupation probability We can apply Bayes theorem: P( j x) = p(x j )P( j ) p(x) = p(x j )P( j ) M j=1 p(x j )P( j ) The posterior probabilities P( j x) give the probability that component j was responsible for generating data point x The P( j x)s are called the component occupation probabilities (or sometimes called the responsibilities) Since they are posterior probabilities: M P( j x) = 1 j=1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 37

51 Parameter estimation If we knew which mixture component was responsible for a data point: we would be able to assign each point unambiguously to a mixture component and we could estimate the mean for each component Gaussian as the sample mean (just like k-means clustering) and we could estimate the covariance as the sample covariance But we don t know which mixture component a data point comes from... Maybe we could use the component occupation probabilities P( j x)? ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 38

52 Gaussian mixture model The most important mixture model is the Gaussian Mixture Model (GMM), where the component densities are Gaussians Consider a GMM, where each component Gaussian N j (x; µ j, Σ j ) has mean µ j and a spherical covariance Σ j = σ 2 j I p(x) = M P( j )p(x j ) = j=1 p(x) M P( j )N j (x; µ j, σj 2 I) j=1 P(1) P(2) P(M) p(x 1) p(x 2) p(x M) x 1 x 2 x d ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 39

53 GMM Parameter estimation when we know which component generated the data Define the indicator variable z jn = 1 if component j generated component x (n) (and 0 otherwise) If z jn wasn t hidden then we could count the number of observed data points generated by j: N N j = n=1 And estimate the mean, variance and mixing parameters as: ˆµ j = ˆσ 2 j = ˆP( j ) = 1 N z jn n z jnx (n) N j n z jn x (n) ˆµ j 2 N j z jn = N j N n ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 40

54 Soft assignment Estimate soft counts based on the component occupation probabilities P( j x (n) ): N Nj = P( j x (n) ) n=1 We can imagine assigning data points to component j weighted by the component occupation probability P( j x (n) ) So we could imagine estimating the mean, variance and prior probabilities as: n ˆµ j = P( j x(n) )x (n) n n P( j = P( j x(n) )x (n) x(n) ) Nj n ˆσ j 2 = P( j x(n) ) x (n) µ j 2 n n P( j = P( j x(n) ) x (n) µ j 2 x(n) ) Nj ˆP( j ) = 1 N n P( j x (n) ) = N j N ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 41

55 EM algorithm Problem! Recall that: P( j x) = p(x j )P( j ) p(x) p(x j )P( j ) = M j=1 p(x j )P( j ) We need to know p(x j ) and P( j ) to estimate the parameters of p( j x), and to estimate P( j )... Solution: an iterative algorithm where each iteration has two parts: Compute the component occupation probabilities P( j x) using the current estimates of the GMM parameters (means, variances, mixing parameters) (E-step) Computer the GMM parameters using the current estimates of the component occupation probabilities (M-step) Starting from some initialization (e.g. using k-means for the means) these steps are alternated until convergence This is called the EM Algorithm and can be shown to maximize the likelihood ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 42

56 Maximum likelihood parameter estimation The likelihood of a data set X = {x (1), x (2),..., x (N) } is given by: N N M L = p(x (n) ) = p(x (n) j )P( j ) n=1 n=1 j=1 We can regard the negative log likelihood as an error function: N E = ln L = ln p(x (n) ) n=1 N M = ln p(x (n) j )P( j ) n=1 j=1 Considering the derivatives of E with respect to the parameters, gives expressions like the previous slide ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 43

57 Example 1 fit using a GMM ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 44

58 Example 1 fit using a GMM Fitted with a two component GMM using EM ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 44

59 Peakily distributed data (Example 2) µ 1 = µ 2 = [0 0] T Σ 1 = 0.1I Σ 2 = 2I ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 45

60 Example 2 fit by a Gaussian µ 1 = µ 2 = [0 0] T Σ 1 = 0.1I Σ 2 = 2I ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 46

61 Example 2 fit by a GMM ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 47

62 Example 2 fit by a GMM Fitted with a two component GMM using EM ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 47

63 Example 2: component Gaussians ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 48

64 Comments on GMMs GMMs trained using the EM algorithm are able to self organize to fit a data set Individual components take responsibility for parts of the data set (probabilistically) Soft assignment to components not hard assignment soft clustering GMMs scale very well, e.g.: large speech recognition systems can have 30,000 GMMs, each with 32 components: sometimes 1 million Gaussian components!! And the parameters all estimated from (a lot of) data by EM ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 49

65 Back to HMMs... P(s 1 s 1 ) P(s 2 s 2 ) P(s 3 s 3 ) s 1 s 2 s 3 s E s I P(s 1 s I ) P(s 2 s 1 ) P(s 3 s 2 ) P(s E s 3 ) p(x s 1 ) p(x s 2 ) p(x s 3 ) x Output distribution: Single multivariate Gaussian with mean µ j, covariance matrix Σ j : b j (x) = p(x S = j ) = N (x; µ j, Σ j ) M-component Gaussian mixture model: M b j (x) = p(x S = j ) = c jm N (x; µ jm, Σ jm ) x x m=1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 50

66 The three problems of HMMs Working with HMMs requires the solution of three problems: 1 Likelihood Determine the overall likelihood of an observation sequence X = (x 1,..., x t,..., x T ) being generated by an HMM. (NB: x t is used to denote x (t) hereafter) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 51

67 The three problems of HMMs Working with HMMs requires the solution of three problems: 1 Likelihood Determine the overall likelihood of an observation sequence X = (x 1,..., x t,..., x T ) being generated by an HMM. (NB: x t is used to denote x (t) hereafter) 2 Decoding Given an observation sequence and an HMM, determine the most probable hidden state sequence ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 51

68 The three problems of HMMs Working with HMMs requires the solution of three problems: 1 Likelihood Determine the overall likelihood of an observation sequence X = (x 1,..., x t,..., x T ) being generated by an HMM. (NB: x t is used to denote x (t) hereafter) 2 Decoding Given an observation sequence and an HMM, determine the most probable hidden state sequence 3 Training Given an observation sequence and an HMM, learn the best HMM parameters λ = {{a jk }, {b j ()}} ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 51

69 1. Likelihood: how to calculate? S 4 33 a34 a a 22 a 11 states a a S 23 S S a 1 01 trellis S time x 2 x 3 x 4 x 5 x 6 x 7 observations x 1 P(X, path j Λ) = P(X path j, Λ)P(path j Λ) = P(X s 0 s 1 s 1 s 1 s 2 s 2 s 3 s 3 s 4, Λ)P(s 0 s 1 s 1 s 1 s 2 s 2 s 3 s 3 s 4 Λ) = b 1 (x 1 )b 1 (x 2 )b 1 (x 3 )b 2 (x 4 )b 2 (x 5 )b 3 (x 6 )b 3 (x 7 )a 01 a 11 a 11 a 12 a 22 a 23 a 33 a 34 P(X Λ) = {path j } P(X, path j Λ) max path j P(X, path j Λ) forward(backward) algorithm Viterbi algorithm ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 52

70 1. Likelihood: The Forward algorithm Goal: determine p(x λ) Sum over all possible state sequences s 1 s 2... s T that could result in the observation sequence X ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 53

71 1. Likelihood: The Forward algorithm Goal: determine p(x λ) Sum over all possible state sequences s 1 s 2... s T that could result in the observation sequence X Rather than enumerating each sequence, compute the probabilities recursively (exploiting the Markov assumption) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 53

72 1. Likelihood: The Forward algorithm Goal: determine p(x λ) Sum over all possible state sequences s 1 s 2... s T that could result in the observation sequence X Rather than enumerating each sequence, compute the probabilities recursively (exploiting the Markov assumption) Hown many paths calculations in p(x λ)? N N N }{{} T times = N T N : number of HMM states T : length of observation e.g. N = 3, T = ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 53

73 1. Likelihood: The Forward algorithm Goal: determine p(x λ) Sum over all possible state sequences s 1 s 2... s T that could result in the observation sequence X Rather than enumerating each sequence, compute the probabilities recursively (exploiting the Markov assumption) Hown many paths calculations in p(x λ)? N N N }{{} T times = N T N : number of HMM states T : length of observation e.g. N = 3, T = Computation complexity for multiplication: O(2T N T ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 53

74 1. Likelihood: The Forward algorithm Goal: determine p(x λ) Sum over all possible state sequences s 1 s 2... s T that could result in the observation sequence X Rather than enumerating each sequence, compute the probabilities recursively (exploiting the Markov assumption) Hown many paths calculations in p(x λ)? N N N }{{} T times = N T N : number of HMM states T : length of observation e.g. N = 3, T = Computation complexity for multiplication: O(2T N T ) The Forward algorithm reduces this to O(TN 2 ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 53

75 Recursive algorithms on HMMs Visualize the problem as a state-time trellis t-1 t t+1 i i i j j j k k k ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 54

76 1. Likelihood: The Forward algorithm Goal: determine p(x λ) Sum over all possible state sequences s 1 s 2... s T that could result in the observation sequence X Rather than enumerating each sequence, compute the probabilities recursively (exploiting the Markov assumption) Forward probability, α t ( j ): the probability of observing the observation sequence x 1... x t and being in state j at time t: α t ( j ) = p(x 1,..., x t, S(t) = j λ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 55

77 1. Likelihood: The Forward recursion Initialization α 0 (s I ) = 1 α 0 ( j ) = 0 if j s I ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 56

78 1. Likelihood: The Forward recursion Initialization α 0 (s I ) = 1 α 0 ( j ) = 0 if j s I Recursion α t ( j ) = N α t 1 ( i )a ij b j (x t ) i=1 1 j N, 1 t T ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 56

79 1. Likelihood: The Forward recursion Initialization α 0 (s I ) = 1 α 0 ( j ) = 0 if j s I Recursion α t ( j ) = Termination N α t 1 ( i )a ij b j (x t ) i=1 p(x λ) = α T (s E ) = 1 j N, 1 t T N α T ( i )a ie i=1 s I : initial state, s E : final state ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 56

80 1. Likelihood: Forward Recursion α t ( j ) = p(x 1,..., x t, S(t) = j λ) = i α ( i ) t 1 j α ( j ) t 1 t 1 a ij a a jj kj Σ j α ( j ) t t i N α t 1 ( i )a ij b j (x t ) i=1 t+1 i j k α ( k ) t 1 k k ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 57

81 Viterbi approximation Instead of summing over all possible state sequences, just consider the most likely ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 58

82 Viterbi approximation Instead of summing over all possible state sequences, just consider the most likely Achieve this by changing the summation to a maximisation in the recursion: V t ( j ) = max V t 1 ( i )a ij b j (x t ) i ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 58

83 Viterbi approximation Instead of summing over all possible state sequences, just consider the most likely Achieve this by changing the summation to a maximisation in the recursion: V t ( j ) = max V t 1 ( i )a ij b j (x t ) i Changing the recursion in this way gives the likelihood of the most probable path ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 58

84 Viterbi approximation Instead of summing over all possible state sequences, just consider the most likely Achieve this by changing the summation to a maximisation in the recursion: V t ( j ) = max V t 1 ( i )a ij b j (x t ) i Changing the recursion in this way gives the likelihood of the most probable path We need to keep track of the states that make up this path by keeping a sequence of backpointers to enable a Viterbi backtrace: the backpointer for each state at each time indicates the previous state on the most probable path ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 58

85 Viterbi Recursion V t ( j ) = max V t 1 ( i )a ij b j (x t ) i Likelihood of the most probable path i V ( i ) t 1 j V ( j ) t 1 t 1 a ij ajj max j akj V ( j ) t t t+1 i i j k V ( k ) t 1 k k ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 59

86 Viterbi Recursion Backpointers to the previous state on the most probable path i i a V ij t 1( i ) bt t ( j )= i ajj j j V ( j ) akj V ( j ) t 1 t 1 t t+1 t i j k V ( k ) t 1 k k ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 60

87 2. Decoding: The Viterbi algorithm Initialization V 0 ( i ) = 1 V 0 ( j ) = 0 bt 0 ( j ) = 0 if j i ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 61

88 2. Decoding: The Viterbi algorithm Initialization V 0 ( i ) = 1 V 0 ( j ) = 0 bt 0 ( j ) = 0 if j i Recursion V t ( j ) = N max i=1 V t 1( i )a ij b j (x t ) bt t ( j ) = arg N max i=1 V t 1( i )a ij b j (x t ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 61

89 2. Decoding: The Viterbi algorithm Initialization V 0 ( i ) = 1 V 0 ( j ) = 0 bt 0 ( j ) = 0 if j i Recursion Termination V t ( j ) = N max i=1 V t 1( i )a ij b j (x t ) bt t ( j ) = arg N max i=1 V t 1( i )a ij b j (x t ) P = V T (s E ) = N max i=1 V T ( i )a ie s T = bt T (q E ) = arg N max i=1 V T ( i )a ie ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 61

90 Viterbi Backtrace Backtrace to find the state sequence of the most probable path t 1 t bt t ( i )= j t+1 i i i V ( i ) t 1 j V ( j ) t 1 j j k V ( k ) t 1 k k bt ( k )= i t+1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 62

91 3. Training: Forward-Backward algorithm Goal: Efficiently estimate the parameters of an HMM λ from an observation sequence ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 63

92 3. Training: Forward-Backward algorithm Goal: Efficiently estimate the parameters of an HMM λ from an observation sequence Assume single Gaussian output probability distribution b j (x) = p(x j ) = N (x; µ j, Σ j ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 63

93 3. Training: Forward-Backward algorithm Goal: Efficiently estimate the parameters of an HMM λ from an observation sequence Assume single Gaussian output probability distribution Parameters λ: Transition probabilities a ij : b j (x) = p(x j ) = N (x; µ j, Σ j ) a ij = 1 Gaussian parameters for state j : mean vector µ j ; covariance matrix Σ j j ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 63

94 Viterbi Training If we knew the state-time alignment, then each observation feature vector could be assigned to a specific state ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 64

95 Viterbi Training If we knew the state-time alignment, then each observation feature vector could be assigned to a specific state A state-time alignment can be obtained using the most probable path obtained by Viterbi decoding ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 64

96 Viterbi Training If we knew the state-time alignment, then each observation feature vector could be assigned to a specific state A state-time alignment can be obtained using the most probable path obtained by Viterbi decoding Maximum likelihood estimate of a ij, if C( i j ) is the count of transitions from i to j â ij = C( i j ) k C( i k ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 64

97 Viterbi Training If we knew the state-time alignment, then each observation feature vector could be assigned to a specific state A state-time alignment can be obtained using the most probable path obtained by Viterbi decoding Maximum likelihood estimate of a ij, if C( i j ) is the count of transitions from i to j â ij = C( i j ) k C( i k ) Likewise if Z j is the set of observed acoustic feature vectors assigned to state j, we can use the standard maximum likelihood estimates for the mean and the covariance: x Z ˆµ j = j x Z j x Z ˆΣ j = j (x ˆµ j )(x ˆµ j ) T Z j ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 64

98 EM Algorithm Viterbi training is an approximation we would like to consider all possible paths ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 65

99 EM Algorithm Viterbi training is an approximation we would like to consider all possible paths In this case rather than having a hard state-time alignment we estimate a probability ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 65

100 EM Algorithm Viterbi training is an approximation we would like to consider all possible paths In this case rather than having a hard state-time alignment we estimate a probability State occupation probability: The probability γ t ( j ) of occupying state j at time t given the sequence of observations. Compare with component occupation probability in a GMM ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 65

101 EM Algorithm Viterbi training is an approximation we would like to consider all possible paths In this case rather than having a hard state-time alignment we estimate a probability State occupation probability: The probability γ t ( j ) of occupying state j at time t given the sequence of observations. Compare with component occupation probability in a GMM We can use this for an iterative algorithm for HMM training: the EM algorithm (whose adaption to HMM is called Baum-Welch algorithm ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 65

102 EM Algorithm Viterbi training is an approximation we would like to consider all possible paths In this case rather than having a hard state-time alignment we estimate a probability State occupation probability: The probability γ t ( j ) of occupying state j at time t given the sequence of observations. Compare with component occupation probability in a GMM We can use this for an iterative algorithm for HMM training: the EM algorithm (whose adaption to HMM is called Baum-Welch algorithm ) Each iteration has two steps: ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 65

103 EM Algorithm Viterbi training is an approximation we would like to consider all possible paths In this case rather than having a hard state-time alignment we estimate a probability State occupation probability: The probability γ t ( j ) of occupying state j at time t given the sequence of observations. Compare with component occupation probability in a GMM We can use this for an iterative algorithm for HMM training: the EM algorithm (whose adaption to HMM is called Baum-Welch algorithm ) Each iteration has two steps: E-step estimate the state occupation probabilities (Expectation) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 65

104 EM Algorithm Viterbi training is an approximation we would like to consider all possible paths In this case rather than having a hard state-time alignment we estimate a probability State occupation probability: The probability γ t ( j ) of occupying state j at time t given the sequence of observations. Compare with component occupation probability in a GMM We can use this for an iterative algorithm for HMM training: the EM algorithm (whose adaption to HMM is called Baum-Welch algorithm ) Each iteration has two steps: E-step estimate the state occupation probabilities (Expectation) M-step re-estimate the HMM parameters based on the estimated state occupation probabilities (Maximisation) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 65

105 Backward probabilities To estimate the state occupation probabilities it is useful to define (recursively) another set of probabilities the Backward probabilities β t ( j ) = p(x t+1,..., x T S(t) = j, λ) The probability of future observations given a the HMM is in state j at time t ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 66

106 Backward probabilities To estimate the state occupation probabilities it is useful to define (recursively) another set of probabilities the Backward probabilities β t ( j ) = p(x t+1,..., x T S(t) = j, λ) The probability of future observations given a the HMM is in state j at time t These can be recursively computed (going backwards in time) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 66

107 Backward probabilities To estimate the state occupation probabilities it is useful to define (recursively) another set of probabilities the Backward probabilities β t ( j ) = p(x t+1,..., x T S(t) = j, λ) The probability of future observations given a the HMM is in state j at time t These can be recursively computed (going backwards in time) Initialisation β T ( i ) = a ie ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 66

108 Backward probabilities To estimate the state occupation probabilities it is useful to define (recursively) another set of probabilities the Backward probabilities β t ( j ) = p(x t+1,..., x T S(t) = j, λ) The probability of future observations given a the HMM is in state j at time t These can be recursively computed (going backwards in time) Initialisation β T ( i ) = a ie Recursion β t ( i ) = N a ij b j (x t+1 )β t+1 ( j ) for t = T 1,..., 1 j=1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 66

109 Backward probabilities To estimate the state occupation probabilities it is useful to define (recursively) another set of probabilities the Backward probabilities β t ( j ) = p(x t+1,..., x T S(t) = j, λ) The probability of future observations given a the HMM is in state j at time t These can be recursively computed (going backwards in time) Initialisation β T ( i ) = a ie Recursion β t ( i ) = Termination N a ij b j (x t+1 )β t+1 ( j ) for t = T 1,..., 1 j=1 p(x λ) = β 0 ( I ) = N a Ij b j (x 1 )β 1 ( j ) = α T (s E ) j=1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 66

110 Backward Recursion N β t ( j ) = p(x t+1,..., x T S(t) = j, λ) = a ij b j (x t+1 )β t+1 ( j ) t 1 i j j β ( j ) t t i Σ a a a ji jk jj j=1 t+1 i β ( i ) t+1 j β ( j ) t+1 k k k β ( k) t+1 ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 67

111 State Occupation Probability The state occupation probability γ t ( j ) is the probability of occupying state j at time t given the sequence of observations ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 68

112 State Occupation Probability The state occupation probability γ t ( j ) is the probability of occupying state j at time t given the sequence of observations Express in terms of the forward and backward probabilities: 1 γ t ( j ) = P(S(t) = j X, λ) = α T (s E ) α t( j )β t ( j ) recalling that p(x λ) = α T (s E ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 68

113 State Occupation Probability The state occupation probability γ t ( j ) is the probability of occupying state j at time t given the sequence of observations Express in terms of the forward and backward probabilities: 1 γ t ( j ) = P(S(t) = j X, λ) = α T (s E ) α t( j )β t ( j ) recalling that p(x λ) = α T (s E ) Since α t ( j )β t ( j ) = p(x 1,..., x t, S(t) = j λ) p(x t+1,..., x T S(t) = j, λ) = p(x 1,..., x t, x t+1,..., x T, S(t) = j λ) = p(x, S(t) = j λ) P(S(t) = j X, λ) = p(x, S(t) = j λ) p(x λ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 68

114 Re-estimation of Gaussian parameters The sum of state occupation probabilities through time for a state, may be regarded as a soft count ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 69

115 Re-estimation of Gaussian parameters The sum of state occupation probabilities through time for a state, may be regarded as a soft count We can use this soft alignment to re-estimate the HMM parameters: ˆµ j = ˆΣ j = T t=1 γ t( j )x t T t=1 γ t( j ) T t=1 γ t( j )(x t ˆµ j )(x ˆµ j ) T T t=1 γ t( j ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 69

116 Re-estimation of transition probabilities Similarly to the state occupation probability, we can estimate ξ t ( i, j ), the probability of being in i at time t and j at t + 1, given the observations: ξ t ( i, j ) = P(S(t) = i, S(t + 1) = j X, λ) P(S(t) = i, S(t + 1) = j, X λ) = p(x Λ) = α t( i )a ij b j (x t+1 )β t+1 ( j ) α T (s E ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 70

117 Re-estimation of transition probabilities Similarly to the state occupation probability, we can estimate ξ t ( i, j ), the probability of being in i at time t and j at t + 1, given the observations: ξ t ( i, j ) = P(S(t) = i, S(t + 1) = j X, λ) P(S(t) = i, S(t + 1) = j, X λ) = p(x Λ) = α t( i )a ij b j (x t+1 )β t+1 ( j ) α T (s E ) We can use this to re-estimate the transition probabilities â ij = T t=1 ξ t( i, j ) N T k=1 t=1 ξ t( i, k ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 70

118 Pulling it all together Iterative estimation of HMM parameters using the EM algorithm. At each iteration E step For all time-state pairs 1 Recursively compute the forward probabilities α t ( j ) and backward probabilities β t ( j ) 2 Compute the state occupation probabilities γ t ( j ) and ξ t ( i, j ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 71

119 Pulling it all together Iterative estimation of HMM parameters using the EM algorithm. At each iteration E step For all time-state pairs 1 Recursively compute the forward probabilities α t ( j ) and backward probabilities β t ( j ) 2 Compute the state occupation probabilities γ t ( j ) and ξ t ( i, j ) M step Based on the estimated state occupation probabilities re-estimate the HMM parameters: mean vectors µ j, covariance matrices Σ j and transition probabilities a ij ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 71

120 Pulling it all together Iterative estimation of HMM parameters using the EM algorithm. At each iteration E step For all time-state pairs 1 Recursively compute the forward probabilities α t ( j ) and backward probabilities β t ( j ) 2 Compute the state occupation probabilities γ t ( j ) and ξ t ( i, j ) M step Based on the estimated state occupation probabilities re-estimate the HMM parameters: mean vectors µ j, covariance matrices Σ j and transition probabilities a ij The application of the EM algorithm to HMM training is sometimes called the Forward-Backward algorithm ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 71

121 Extension to a corpus of utterances We usually train from a large corpus of R utterances If x r t is the t th frame of the r th utterance X r then we can compute the probabilities α r t( j ), β r t ( j ), γ r t ( j ) and ξ r t ( i, j ) as before The re-estimates are as before, except we must sum over the R utterances, eg: R T r=1 t=1 ˆµ j = γr t ( j )x r t R T r=1 t=1 γr t ( j ) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 72

122 Extension to Gaussian mixture model (GMM) The assumption of a Gaussian distribution at each state is very strong; in practice the acoustic feature vectors associated with a state may be strongly non-gaussian ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 73

123 Extension to Gaussian mixture model (GMM) The assumption of a Gaussian distribution at each state is very strong; in practice the acoustic feature vectors associated with a state may be strongly non-gaussian In this case an M-component Gaussian mixture model is an appropriate density function: b j (x) = p(x j ) = M c jm N (x; µ jm, Σ jm ) m=1 Given enough components, this family of functions can model any distribution. ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 73

124 Extension to Gaussian mixture model (GMM) The assumption of a Gaussian distribution at each state is very strong; in practice the acoustic feature vectors associated with a state may be strongly non-gaussian In this case an M-component Gaussian mixture model is an appropriate density function: b j (x) = p(x j ) = M c jm N (x; µ jm, Σ jm ) m=1 Given enough components, this family of functions can model any distribution. Train using the EM algorithm, in which the component estimation probabilities are estimated in the E-step ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 73

125 EM training of HMM/GMM Rather than estimating the state-time alignment, we estimate the component/state-time alignment, and component-state occupation probabilities γ t ( j, m): the probability of occupying mixture component m of state j at time t ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 74

126 EM training of HMM/GMM Rather than estimating the state-time alignment, we estimate the component/state-time alignment, and component-state occupation probabilities γ t ( j, m): the probability of occupying mixture component m of state j at time t We can thus re-estimate the mean of mixture component m of state j as follows ˆµ jm = T t=1 γ t( j, m)x t T t=1 γ t( j, m) And likewise for the covariance matrices (mixture models often use diagonal covariance matrices) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 74

127 EM training of HMM/GMM Rather than estimating the state-time alignment, we estimate the component/state-time alignment, and component-state occupation probabilities γ t ( j, m): the probability of occupying mixture component m of state j at time t We can thus re-estimate the mean of mixture component m of state j as follows ˆµ jm = T t=1 γ t( j, m)x t T t=1 γ t( j, m) And likewise for the covariance matrices (mixture models often use diagonal covariance matrices) The mixture coefficients are re-estimated in a similar way to transition probabilities: T t=1 ĉ jm = γ t( j, m) M T l=1 t=1 γ t( j, l) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 74

128 Doing the computation The forward, backward and Viterbi recursions result in a long sequence of probabilities being multiplied This can cause floating point underflow problems In practice computations are performed in the log domain (in which multiplies become adds) Working in the log domain also avoids needing to perform the exponentiation when computing Gaussians ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 75

129 A note on HMM topology left to right model a 11 a a 22 a a 33 parallel path left to right model a 11 a 12 a a 22 a 23 a a 33 a 34 a a 44 a a 55 ergodic model a 11 a 12 a 13 a 14 a 15 a 21 a 22 a 23 a 24 a 25 a 31 a 32 a 33 a 34 a 35 a 41 a 42 a 43 a 44 a 45 a 51 a 52 a 53 a 54 a 55 Speech recognition: Speaker recognition: left-to-right HMM with 3 5 states ergodic HMM ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 76

130 A note on HMM emission probabilities a 11 a 22 a 33 1 a 01 b (x) a 12 b (x) 2 a b (x) 3 a emission pdfs Emission prob. Continuous (density) HMM continuous density GMM, NN/DNN Discrete (probability) HMM discrete probability VQ Semi-continuous HMM continuous density tied mixture (tied-mixture HMM) ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 77

131 Summary: HMMs HMMs provide a generative model for statistical speech recognition Three key problems 1 Computing the overall likelihood: the Forward algorithm 2 Decoding the most likely state sequence: the Viterbi algorithm 3 Estimating the most likely parameters: the EM (Forward-Backward) algorithm Solutions to these problems are tractable due to the two key HMM assumptions 1 Conditional independence of observations given the current state 2 Markov assumption on the states ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 78

132 References: HMMs Gales and Young (2007). The Application of Hidden Markov Models in Speech Recognition, Foundations and Trends in Signal Processing, 1 (3), : section 2.2. Jurafsky and Martin (2008). Speech and Language Processing (2nd ed.): sections ; 9.2; 9.4. (Errata at SLP2-PIEV-Errata.html) Rabiner and Juang (1989). An introduction to hidden Markov models, IEEE ASSP Magazine, 3 (1), Renals and Hain (2010). Speech Recognition, Computational Linguistics and Natural Language Processing Handbook, Clark, Fox and Lappin (eds.), Blackwells. ASR Lectures 4&5 Hidden Markov Models and Gaussian Mixture Models 79

Hidden Markov Models and Gaussian Mixture Models

Hidden Markov Models and Gaussian Mixture Models Hidden Markov Models and Gaussian Mixture Models Hiroshi Shimodaira and Steve Renals Automatic Speech Recognition ASR Lectures 4&5 25&29 January 2018 ASR Lectures 4&5 Hidden Markov Models and Gaussian

More information

University of Cambridge. MPhil in Computer Speech Text & Internet Technology. Module: Speech Processing II. Lecture 2: Hidden Markov Models I

University of Cambridge. MPhil in Computer Speech Text & Internet Technology. Module: Speech Processing II. Lecture 2: Hidden Markov Models I University of Cambridge MPhil in Computer Speech Text & Internet Technology Module: Speech Processing II Lecture 2: Hidden Markov Models I o o o o o 1 2 3 4 T 1 b 2 () a 12 2 a 3 a 4 5 34 a 23 b () b ()

More information

COM336: Neural Computing

COM336: Neural Computing COM336: Neural Computing http://www.dcs.shef.ac.uk/ sjr/com336/ Lecture 2: Density Estimation Steve Renals Department of Computer Science University of Sheffield Sheffield S1 4DP UK email: s.renals@dcs.shef.ac.uk

More information

Hidden Markov Models

Hidden Markov Models Hidden Markov Models Dr Philip Jackson Centre for Vision, Speech & Signal Processing University of Surrey, UK 1 3 2 http://www.ee.surrey.ac.uk/personal/p.jackson/isspr/ Outline 1. Recognizing patterns

More information

CS 136a Lecture 7 Speech Recognition Architecture: Training models with the Forward backward algorithm

CS 136a Lecture 7 Speech Recognition Architecture: Training models with the Forward backward algorithm + September13, 2016 Professor Meteer CS 136a Lecture 7 Speech Recognition Architecture: Training models with the Forward backward algorithm Thanks to Dan Jurafsky for these slides + ASR components n Feature

More information

Hidden Markov Modelling

Hidden Markov Modelling Hidden Markov Modelling Introduction Problem formulation Forward-Backward algorithm Viterbi search Baum-Welch parameter estimation Other considerations Multiple observation sequences Phone-based models

More information

HIDDEN MARKOV MODELS IN SPEECH RECOGNITION

HIDDEN MARKOV MODELS IN SPEECH RECOGNITION HIDDEN MARKOV MODELS IN SPEECH RECOGNITION Wayne Ward Carnegie Mellon University Pittsburgh, PA 1 Acknowledgements Much of this talk is derived from the paper "An Introduction to Hidden Markov Models",

More information

Speech and Language Processing. Chapter 9 of SLP Automatic Speech Recognition (II)

Speech and Language Processing. Chapter 9 of SLP Automatic Speech Recognition (II) Speech and Language Processing Chapter 9 of SLP Automatic Speech Recognition (II) Outline for ASR ASR Architecture The Noisy Channel Model Five easy pieces of an ASR system 1) Language Model 2) Lexicon/Pronunciation

More information

Lecture 11: Hidden Markov Models

Lecture 11: Hidden Markov Models Lecture 11: Hidden Markov Models Cognitive Systems - Machine Learning Cognitive Systems, Applied Computer Science, Bamberg University slides by Dr. Philip Jackson Centre for Vision, Speech & Signal Processing

More information

Automatic Speech Recognition (CS753)

Automatic Speech Recognition (CS753) Automatic Speech Recognition (CS753) Lecture 6: Hidden Markov Models (Part II) Instructor: Preethi Jyothi Aug 10, 2017 Recall: Computing Likelihood Problem 1 (Likelihood): Given an HMM l =(A, B) and an

More information

STA 414/2104: Machine Learning

STA 414/2104: Machine Learning STA 414/2104: Machine Learning Russ Salakhutdinov Department of Computer Science! Department of Statistics! rsalakhu@cs.toronto.edu! http://www.cs.toronto.edu/~rsalakhu/ Lecture 9 Sequential Data So far

More information

STA 4273H: Statistical Machine Learning

STA 4273H: Statistical Machine Learning STA 4273H: Statistical Machine Learning Russ Salakhutdinov Department of Statistics! rsalakhu@utstat.toronto.edu! http://www.utstat.utoronto.ca/~rsalakhu/ Sidney Smith Hall, Room 6002 Lecture 11 Project

More information

Statistical Sequence Recognition and Training: An Introduction to HMMs

Statistical Sequence Recognition and Training: An Introduction to HMMs Statistical Sequence Recognition and Training: An Introduction to HMMs EECS 225D Nikki Mirghafori nikki@icsi.berkeley.edu March 7, 2005 Credit: many of the HMM slides have been borrowed and adapted, with

More information

Lecture 3: ASR: HMMs, Forward, Viterbi

Lecture 3: ASR: HMMs, Forward, Viterbi Original slides by Dan Jurafsky CS 224S / LINGUIST 285 Spoken Language Processing Andrew Maas Stanford University Spring 2017 Lecture 3: ASR: HMMs, Forward, Viterbi Fun informative read on phonetics The

More information

Parametric Unsupervised Learning Expectation Maximization (EM) Lecture 20.a

Parametric Unsupervised Learning Expectation Maximization (EM) Lecture 20.a Parametric Unsupervised Learning Expectation Maximization (EM) Lecture 20.a Some slides are due to Christopher Bishop Limitations of K-means Hard assignments of data points to clusters small shift of a

More information

EECS E6870: Lecture 4: Hidden Markov Models

EECS E6870: Lecture 4: Hidden Markov Models EECS E6870: Lecture 4: Hidden Markov Models Stanley F. Chen, Michael A. Picheny and Bhuvana Ramabhadran IBM T. J. Watson Research Center Yorktown Heights, NY 10549 stanchen@us.ibm.com, picheny@us.ibm.com,

More information

Hidden Markov Models

Hidden Markov Models Hidden Markov Models Lecture Notes Speech Communication 2, SS 2004 Erhard Rank/Franz Pernkopf Signal Processing and Speech Communication Laboratory Graz University of Technology Inffeldgasse 16c, A-8010

More information

Statistical NLP Spring Digitizing Speech

Statistical NLP Spring Digitizing Speech Statistical NLP Spring 2008 Lecture 10: Acoustic Models Dan Klein UC Berkeley Digitizing Speech 1 Frame Extraction A frame (25 ms wide) extracted every 10 ms 25 ms 10ms... a 1 a 2 a 3 Figure from Simon

More information

Digitizing Speech. Statistical NLP Spring Frame Extraction. Gaussian Emissions. Vector Quantization. HMMs for Continuous Observations? ...

Digitizing Speech. Statistical NLP Spring Frame Extraction. Gaussian Emissions. Vector Quantization. HMMs for Continuous Observations? ... Statistical NLP Spring 2008 Digitizing Speech Lecture 10: Acoustic Models Dan Klein UC Berkeley Frame Extraction A frame (25 ms wide extracted every 10 ms 25 ms 10ms... a 1 a 2 a 3 Figure from Simon Arnfield

More information

Speech Recognition HMM

Speech Recognition HMM Speech Recognition HMM Jan Černocký, Valentina Hubeika {cernocky ihubeika}@fit.vutbr.cz FIT BUT Brno Speech Recognition HMM Jan Černocký, Valentina Hubeika, DCGM FIT BUT Brno 1/38 Agenda Recap variability

More information

Engineering Part IIB: Module 4F11 Speech and Language Processing Lectures 4/5 : Speech Recognition Basics

Engineering Part IIB: Module 4F11 Speech and Language Processing Lectures 4/5 : Speech Recognition Basics Engineering Part IIB: Module 4F11 Speech and Language Processing Lectures 4/5 : Speech Recognition Basics Phil Woodland: pcw@eng.cam.ac.uk Lent 2013 Engineering Part IIB: Module 4F11 What is Speech Recognition?

More information

Hidden Markov Models. Aarti Singh Slides courtesy: Eric Xing. Machine Learning / Nov 8, 2010

Hidden Markov Models. Aarti Singh Slides courtesy: Eric Xing. Machine Learning / Nov 8, 2010 Hidden Markov Models Aarti Singh Slides courtesy: Eric Xing Machine Learning 10-701/15-781 Nov 8, 2010 i.i.d to sequential data So far we assumed independent, identically distributed data Sequential data

More information

Basic math for biology

Basic math for biology Basic math for biology Lei Li Florida State University, Feb 6, 2002 The EM algorithm: setup Parametric models: {P θ }. Data: full data (Y, X); partial data Y. Missing data: X. Likelihood and maximum likelihood

More information

Hidden Markov Models. By Parisa Abedi. Slides courtesy: Eric Xing

Hidden Markov Models. By Parisa Abedi. Slides courtesy: Eric Xing Hidden Markov Models By Parisa Abedi Slides courtesy: Eric Xing i.i.d to sequential data So far we assumed independent, identically distributed data Sequential (non i.i.d.) data Time-series data E.g. Speech

More information

The Noisy Channel Model. Statistical NLP Spring Mel Freq. Cepstral Coefficients. Frame Extraction ... Lecture 10: Acoustic Models

The Noisy Channel Model. Statistical NLP Spring Mel Freq. Cepstral Coefficients. Frame Extraction ... Lecture 10: Acoustic Models Statistical NLP Spring 2009 The Noisy Channel Model Lecture 10: Acoustic Models Dan Klein UC Berkeley Search through space of all possible sentences. Pick the one that is most probable given the waveform.

More information

Statistical NLP Spring The Noisy Channel Model

Statistical NLP Spring The Noisy Channel Model Statistical NLP Spring 2009 Lecture 10: Acoustic Models Dan Klein UC Berkeley The Noisy Channel Model Search through space of all possible sentences. Pick the one that is most probable given the waveform.

More information

Machine Learning for natural language processing

Machine Learning for natural language processing Machine Learning for natural language processing Hidden Markov Models Laura Kallmeyer Heinrich-Heine-Universität Düsseldorf Summer 2016 1 / 33 Introduction So far, we have classified texts/observations

More information

Basic Text Analysis. Hidden Markov Models. Joakim Nivre. Uppsala University Department of Linguistics and Philology

Basic Text Analysis. Hidden Markov Models. Joakim Nivre. Uppsala University Department of Linguistics and Philology Basic Text Analysis Hidden Markov Models Joakim Nivre Uppsala University Department of Linguistics and Philology joakimnivre@lingfiluuse Basic Text Analysis 1(33) Hidden Markov Models Markov models are

More information

CS 136 Lecture 5 Acoustic modeling Phoneme modeling

CS 136 Lecture 5 Acoustic modeling Phoneme modeling + September 9, 2016 Professor Meteer CS 136 Lecture 5 Acoustic modeling Phoneme modeling Thanks to Dan Jurafsky for these slides + Directly Modeling Continuous Observations n Gaussians n Univariate Gaussians

More information

Hidden Markov Models The three basic HMM problems (note: change in notation) Mitch Marcus CSE 391

Hidden Markov Models The three basic HMM problems (note: change in notation) Mitch Marcus CSE 391 Hidden Markov Models The three basic HMM problems (note: change in notation) Mitch Marcus CSE 391 Parameters of an HMM States: A set of states S=s 1, s n Transition probabilities: A= a 1,1, a 1,2,, a n,n

More information

Part of Speech Tagging: Viterbi, Forward, Backward, Forward- Backward, Baum-Welch. COMP-599 Oct 1, 2015

Part of Speech Tagging: Viterbi, Forward, Backward, Forward- Backward, Baum-Welch. COMP-599 Oct 1, 2015 Part of Speech Tagging: Viterbi, Forward, Backward, Forward- Backward, Baum-Welch COMP-599 Oct 1, 2015 Announcements Research skills workshop today 3pm-4:30pm Schulich Library room 313 Start thinking about

More information

Hidden Markov Model and Speech Recognition

Hidden Markov Model and Speech Recognition 1 Dec,2006 Outline Introduction 1 Introduction 2 3 4 5 Introduction What is Speech Recognition? Understanding what is being said Mapping speech data to textual information Speech Recognition is indeed

More information

Introduction to Machine Learning CMU-10701

Introduction to Machine Learning CMU-10701 Introduction to Machine Learning CMU-10701 Hidden Markov Models Barnabás Póczos & Aarti Singh Slides courtesy: Eric Xing i.i.d to sequential data So far we assumed independent, identically distributed

More information

CISC 889 Bioinformatics (Spring 2004) Hidden Markov Models (II)

CISC 889 Bioinformatics (Spring 2004) Hidden Markov Models (II) CISC 889 Bioinformatics (Spring 24) Hidden Markov Models (II) a. Likelihood: forward algorithm b. Decoding: Viterbi algorithm c. Model building: Baum-Welch algorithm Viterbi training Hidden Markov models

More information

Parametric Models Part III: Hidden Markov Models

Parametric Models Part III: Hidden Markov Models Parametric Models Part III: Hidden Markov Models Selim Aksoy Department of Computer Engineering Bilkent University saksoy@cs.bilkent.edu.tr CS 551, Spring 2014 CS 551, Spring 2014 c 2014, Selim Aksoy (Bilkent

More information

Lecture 5: GMM Acoustic Modeling and Feature Extraction

Lecture 5: GMM Acoustic Modeling and Feature Extraction CS 224S / LINGUIST 285 Spoken Language Processing Andrew Maas Stanford University Spring 2017 Lecture 5: GMM Acoustic Modeling and Feature Extraction Original slides by Dan Jurafsky Outline for Today Acoustic

More information

Introduction to Neural Networks

Introduction to Neural Networks Introduction to Neural Networks Steve Renals Automatic Speech Recognition ASR Lecture 10 24 February 2014 ASR Lecture 10 Introduction to Neural Networks 1 Neural networks for speech recognition Introduction

More information

Lecture 3. Gaussian Mixture Models and Introduction to HMM s. Michael Picheny, Bhuvana Ramabhadran, Stanley F. Chen, Markus Nussbaum-Thom

Lecture 3. Gaussian Mixture Models and Introduction to HMM s. Michael Picheny, Bhuvana Ramabhadran, Stanley F. Chen, Markus Nussbaum-Thom Lecture 3 Gaussian Mixture Models and Introduction to HMM s Michael Picheny, Bhuvana Ramabhadran, Stanley F. Chen, Markus Nussbaum-Thom Watson Group IBM T.J. Watson Research Center Yorktown Heights, New

More information

Statistical Methods for NLP

Statistical Methods for NLP Statistical Methods for NLP Information Extraction, Hidden Markov Models Sameer Maskey Week 5, Oct 3, 2012 *many slides provided by Bhuvana Ramabhadran, Stanley Chen, Michael Picheny Speech Recognition

More information

ASR using Hidden Markov Model : A tutorial

ASR using Hidden Markov Model : A tutorial ASR using Hidden Markov Model : A tutorial Samudravijaya K Workshop on ASR @BAMU; 14-OCT-11 samudravijaya@gmail.com Tata Institute of Fundamental Research Samudravijaya K Workshop on ASR @BAMU; 14-OCT-11

More information

Data-Intensive Computing with MapReduce

Data-Intensive Computing with MapReduce Data-Intensive Computing with MapReduce Session 8: Sequence Labeling Jimmy Lin University of Maryland Thursday, March 14, 2013 This work is licensed under a Creative Commons Attribution-Noncommercial-Share

More information

Hidden Markov Models Part 2: Algorithms

Hidden Markov Models Part 2: Algorithms Hidden Markov Models Part 2: Algorithms CSE 6363 Machine Learning Vassilis Athitsos Computer Science and Engineering Department University of Texas at Arlington 1 Hidden Markov Model An HMM consists of:

More information

Inf2b Learning and Data

Inf2b Learning and Data Inf2b Learning and Data Lecture 13: Review (Credit: Hiroshi Shimodaira Iain Murray and Steve Renals) Centre for Speech Technology Research (CSTR) School of Informatics University of Edinburgh http://www.inf.ed.ac.uk/teaching/courses/inf2b/

More information

Hidden Markov Model. Ying Wu. Electrical Engineering and Computer Science Northwestern University Evanston, IL 60208

Hidden Markov Model. Ying Wu. Electrical Engineering and Computer Science Northwestern University Evanston, IL 60208 Hidden Markov Model Ying Wu Electrical Engineering and Computer Science Northwestern University Evanston, IL 60208 http://www.eecs.northwestern.edu/~yingwu 1/19 Outline Example: Hidden Coin Tossing Hidden

More information

Hidden Markov Models Hamid R. Rabiee

Hidden Markov Models Hamid R. Rabiee Hidden Markov Models Hamid R. Rabiee 1 Hidden Markov Models (HMMs) In the previous slides, we have seen that in many cases the underlying behavior of nature could be modeled as a Markov process. However

More information

Cheng Soon Ong & Christian Walder. Canberra February June 2018

Cheng Soon Ong & Christian Walder. Canberra February June 2018 Cheng Soon Ong & Christian Walder Research Group and College of Engineering and Computer Science Canberra February June 2018 Outlines Overview Introduction Linear Algebra Probability Linear Regression

More information

The Noisy Channel Model. CS 294-5: Statistical Natural Language Processing. Speech Recognition Architecture. Digitizing Speech

The Noisy Channel Model. CS 294-5: Statistical Natural Language Processing. Speech Recognition Architecture. Digitizing Speech CS 294-5: Statistical Natural Language Processing The Noisy Channel Model Speech Recognition II Lecture 21: 11/29/05 Search through space of all possible sentences. Pick the one that is most probable given

More information

Automatic Speech Recognition (CS753)

Automatic Speech Recognition (CS753) Automatic Speech Recognition (CS753) Lecture 21: Speaker Adaptation Instructor: Preethi Jyothi Oct 23, 2017 Speaker variations Major cause of variability in speech is the differences between speakers Speaking

More information

An Evolutionary Programming Based Algorithm for HMM training

An Evolutionary Programming Based Algorithm for HMM training An Evolutionary Programming Based Algorithm for HMM training Ewa Figielska,Wlodzimierz Kasprzak Institute of Control and Computation Engineering, Warsaw University of Technology ul. Nowowiejska 15/19,

More information

Machine Learning. Gaussian Mixture Models. Zhiyao Duan & Bryan Pardo, Machine Learning: EECS 349 Fall

Machine Learning. Gaussian Mixture Models. Zhiyao Duan & Bryan Pardo, Machine Learning: EECS 349 Fall Machine Learning Gaussian Mixture Models Zhiyao Duan & Bryan Pardo, Machine Learning: EECS 349 Fall 2012 1 The Generative Model POV We think of the data as being generated from some process. We assume

More information

Sequence labeling. Taking collective a set of interrelated instances x 1,, x T and jointly labeling them

Sequence labeling. Taking collective a set of interrelated instances x 1,, x T and jointly labeling them HMM, MEMM and CRF 40-957 Special opics in Artificial Intelligence: Probabilistic Graphical Models Sharif University of echnology Soleymani Spring 2014 Sequence labeling aking collective a set of interrelated

More information

A Gentle Tutorial of the EM Algorithm and its Application to Parameter Estimation for Gaussian Mixture and Hidden Markov Models

A Gentle Tutorial of the EM Algorithm and its Application to Parameter Estimation for Gaussian Mixture and Hidden Markov Models A Gentle Tutorial of the EM Algorithm and its Application to Parameter Estimation for Gaussian Mixture and Hidden Markov Models Jeff A. Bilmes (bilmes@cs.berkeley.edu) International Computer Science Institute

More information

Temporal Modeling and Basic Speech Recognition

Temporal Modeling and Basic Speech Recognition UNIVERSITY ILLINOIS @ URBANA-CHAMPAIGN OF CS 498PS Audio Computing Lab Temporal Modeling and Basic Speech Recognition Paris Smaragdis paris@illinois.edu paris.cs.illinois.edu Today s lecture Recognizing

More information

Expectation Maximization (EM)

Expectation Maximization (EM) Expectation Maximization (EM) The Expectation Maximization (EM) algorithm is one approach to unsupervised, semi-supervised, or lightly supervised learning. In this kind of learning either no labels are

More information

10. Hidden Markov Models (HMM) for Speech Processing. (some slides taken from Glass and Zue course)

10. Hidden Markov Models (HMM) for Speech Processing. (some slides taken from Glass and Zue course) 10. Hidden Markov Models (HMM) for Speech Processing (some slides taken from Glass and Zue course) Definition of an HMM The HMM are powerful statistical methods to characterize the observed samples of

More information

Lecture 4: Hidden Markov Models: An Introduction to Dynamic Decision Making. November 11, 2010

Lecture 4: Hidden Markov Models: An Introduction to Dynamic Decision Making. November 11, 2010 Hidden Lecture 4: Hidden : An Introduction to Dynamic Decision Making November 11, 2010 Special Meeting 1/26 Markov Model Hidden When a dynamical system is probabilistic it may be determined by the transition

More information

Recap: HMM. ANLP Lecture 9: Algorithms for HMMs. More general notation. Recap: HMM. Elements of HMM: Sharon Goldwater 4 Oct 2018.

Recap: HMM. ANLP Lecture 9: Algorithms for HMMs. More general notation. Recap: HMM. Elements of HMM: Sharon Goldwater 4 Oct 2018. Recap: HMM ANLP Lecture 9: Algorithms for HMMs Sharon Goldwater 4 Oct 2018 Elements of HMM: Set of states (tags) Output alphabet (word types) Start state (beginning of sentence) State transition probabilities

More information

Hidden Markov Models

Hidden Markov Models CS769 Spring 2010 Advanced Natural Language Processing Hidden Markov Models Lecturer: Xiaojin Zhu jerryzhu@cs.wisc.edu 1 Part-of-Speech Tagging The goal of Part-of-Speech (POS) tagging is to label each

More information

Robert Collins CSE586 CSE 586, Spring 2015 Computer Vision II

Robert Collins CSE586 CSE 586, Spring 2015 Computer Vision II CSE 586, Spring 2015 Computer Vision II Hidden Markov Model and Kalman Filter Recall: Modeling Time Series State-Space Model: You have a Markov chain of latent (unobserved) states Each state generates

More information

The Noisy Channel Model. Statistical NLP Spring Mel Freq. Cepstral Coefficients. Frame Extraction ... Lecture 9: Acoustic Models

The Noisy Channel Model. Statistical NLP Spring Mel Freq. Cepstral Coefficients. Frame Extraction ... Lecture 9: Acoustic Models Statistical NLP Spring 2010 The Noisy Channel Model Lecture 9: Acoustic Models Dan Klein UC Berkeley Acoustic model: HMMs over word positions with mixtures of Gaussians as emissions Language model: Distributions

More information

Statistical Pattern Recognition

Statistical Pattern Recognition Statistical Pattern Recognition Expectation Maximization (EM) and Mixture Models Hamid R. Rabiee Jafar Muhammadi, Mohammad J. Hosseini Spring 2014 http://ce.sharif.edu/courses/92-93/2/ce725-2 Agenda Expectation-maximization

More information

Outline of Today s Lecture

Outline of Today s Lecture University of Washington Department of Electrical Engineering Computer Speech Processing EE516 Winter 2005 Jeff A. Bilmes Lecture 12 Slides Feb 23 rd, 2005 Outline of Today s

More information

Linear Dynamical Systems (Kalman filter)

Linear Dynamical Systems (Kalman filter) Linear Dynamical Systems (Kalman filter) (a) Overview of HMMs (b) From HMMs to Linear Dynamical Systems (LDS) 1 Markov Chains with Discrete Random Variables x 1 x 2 x 3 x T Let s assume we have discrete

More information

Statistical Processing of Natural Language

Statistical Processing of Natural Language Statistical Processing of Natural Language and DMKM - Universitat Politècnica de Catalunya and 1 2 and 3 1. Observation Probability 2. Best State Sequence 3. Parameter Estimation 4 Graphical and Generative

More information

CMSC 723: Computational Linguistics I Session #5 Hidden Markov Models. The ischool University of Maryland. Wednesday, September 30, 2009

CMSC 723: Computational Linguistics I Session #5 Hidden Markov Models. The ischool University of Maryland. Wednesday, September 30, 2009 CMSC 723: Computational Linguistics I Session #5 Hidden Markov Models Jimmy Lin The ischool University of Maryland Wednesday, September 30, 2009 Today s Agenda The great leap forward in NLP Hidden Markov

More information

Note Set 5: Hidden Markov Models

Note Set 5: Hidden Markov Models Note Set 5: Hidden Markov Models Probabilistic Learning: Theory and Algorithms, CS 274A, Winter 2016 1 Hidden Markov Models (HMMs) 1.1 Introduction Consider observed data vectors x t that are d-dimensional

More information

Brief Introduction of Machine Learning Techniques for Content Analysis

Brief Introduction of Machine Learning Techniques for Content Analysis 1 Brief Introduction of Machine Learning Techniques for Content Analysis Wei-Ta Chu 2008/11/20 Outline 2 Overview Gaussian Mixture Model (GMM) Hidden Markov Model (HMM) Support Vector Machine (SVM) Overview

More information

p(d θ ) l(θ ) 1.2 x x x

p(d θ ) l(θ ) 1.2 x x x p(d θ ).2 x 0-7 0.8 x 0-7 0.4 x 0-7 l(θ ) -20-40 -60-80 -00 2 3 4 5 6 7 θ ˆ 2 3 4 5 6 7 θ ˆ 2 3 4 5 6 7 θ θ x FIGURE 3.. The top graph shows several training points in one dimension, known or assumed to

More information

Lecture 12: Algorithms for HMMs

Lecture 12: Algorithms for HMMs Lecture 12: Algorithms for HMMs Nathan Schneider (some slides from Sharon Goldwater; thanks to Jonathan May for bug fixes) ENLP 26 February 2018 Recap: tagging POS tagging is a sequence labelling task.

More information

Markov Chains and Hidden Markov Models

Markov Chains and Hidden Markov Models Chapter 1 Markov Chains and Hidden Markov Models In this chapter, we will introduce the concept of Markov chains, and show how Markov chains can be used to model signals using structures such as hidden

More information

Shankar Shivappa University of California, San Diego April 26, CSE 254 Seminar in learning algorithms

Shankar Shivappa University of California, San Diego April 26, CSE 254 Seminar in learning algorithms Recognition of Visual Speech Elements Using Adaptively Boosted Hidden Markov Models. Say Wei Foo, Yong Lian, Liang Dong. IEEE Transactions on Circuits and Systems for Video Technology, May 2004. Shankar

More information

The Expectation-Maximization Algorithm

The Expectation-Maximization Algorithm 1/29 EM & Latent Variable Models Gaussian Mixture Models EM Theory The Expectation-Maximization Algorithm Mihaela van der Schaar Department of Engineering Science University of Oxford MLE for Latent Variable

More information

We Live in Exciting Times. CSCI-567: Machine Learning (Spring 2019) Outline. Outline. ACM (an international computing research society) has named

We Live in Exciting Times. CSCI-567: Machine Learning (Spring 2019) Outline. Outline. ACM (an international computing research society) has named We Live in Exciting Times ACM (an international computing research society) has named CSCI-567: Machine Learning (Spring 2019) Prof. Victor Adamchik U of Southern California Apr. 2, 2019 Yoshua Bengio,

More information

Linear Dynamical Systems

Linear Dynamical Systems Linear Dynamical Systems Sargur N. srihari@cedar.buffalo.edu Machine Learning Course: http://www.cedar.buffalo.edu/~srihari/cse574/index.html Two Models Described by Same Graph Latent variables Observations

More information

Hidden Markov Models in Language Processing

Hidden Markov Models in Language Processing Hidden Markov Models in Language Processing Dustin Hillard Lecture notes courtesy of Prof. Mari Ostendorf Outline Review of Markov models What is an HMM? Examples General idea of hidden variables: implications

More information

Lecture 12: Algorithms for HMMs

Lecture 12: Algorithms for HMMs Lecture 12: Algorithms for HMMs Nathan Schneider (some slides from Sharon Goldwater; thanks to Jonathan May for bug fixes) ENLP 17 October 2016 updated 9 September 2017 Recap: tagging POS tagging is a

More information

Weighted Finite-State Transducers in Computational Biology

Weighted Finite-State Transducers in Computational Biology Weighted Finite-State Transducers in Computational Biology Mehryar Mohri Courant Institute of Mathematical Sciences mohri@cims.nyu.edu Joint work with Corinna Cortes (Google Research). 1 This Tutorial

More information

1. Markov models. 1.1 Markov-chain

1. Markov models. 1.1 Markov-chain 1. Markov models 1.1 Markov-chain Let X be a random variable X = (X 1,..., X t ) taking values in some set S = {s 1,..., s N }. The sequence is Markov chain if it has the following properties: 1. Limited

More information

Hidden Markov Models

Hidden Markov Models Hidden Markov Models CI/CI(CS) UE, SS 2015 Christian Knoll Signal Processing and Speech Communication Laboratory Graz University of Technology June 23, 2015 CI/CI(CS) SS 2015 June 23, 2015 Slide 1/26 Content

More information

Introduction to Machine Learning. Maximum Likelihood and Bayesian Inference. Lecturers: Eran Halperin, Lior Wolf

Introduction to Machine Learning. Maximum Likelihood and Bayesian Inference. Lecturers: Eran Halperin, Lior Wolf 1 Introduction to Machine Learning Maximum Likelihood and Bayesian Inference Lecturers: Eran Halperin, Lior Wolf 2014-15 We know that X ~ B(n,p), but we do not know p. We get a random sample from X, a

More information

Expectation Maximization (EM)

Expectation Maximization (EM) Expectation Maximization (EM) The EM algorithm is used to train models involving latent variables using training data in which the latent variables are not observed (unlabeled data). This is to be contrasted

More information

Gaussians. Hiroshi Shimodaira. January-March Continuous random variables

Gaussians. Hiroshi Shimodaira. January-March Continuous random variables Cumulative Distribution Function Gaussians Hiroshi Shimodaira January-March 9 In this chapter we introduce the basics of how to build probabilistic models of continuous-valued data, including the most

More information

HMM part 1. Dr Philip Jackson

HMM part 1. Dr Philip Jackson Centre for Vision Speech & Signal Processing University of Surrey, Guildford GU2 7XH. HMM part 1 Dr Philip Jackson Probability fundamentals Markov models State topology diagrams Hidden Markov models -

More information

L23: hidden Markov models

L23: hidden Markov models L23: hidden Markov models Discrete Markov processes Hidden Markov models Forward and Backward procedures The Viterbi algorithm This lecture is based on [Rabiner and Juang, 1993] Introduction to Speech

More information

order is number of previous outputs

order is number of previous outputs Markov Models Lecture : Markov and Hidden Markov Models PSfrag Use past replacements as state. Next output depends on previous output(s): y t = f[y t, y t,...] order is number of previous outputs y t y

More information

Expectation Maximization

Expectation Maximization Expectation Maximization Aaron C. Courville Université de Montréal Note: Material for the slides is taken directly from a presentation prepared by Christopher M. Bishop Learning in DAGs Two things could

More information

Hidden Markov Models. Dr. Naomi Harte

Hidden Markov Models. Dr. Naomi Harte Hidden Markov Models Dr. Naomi Harte The Talk Hidden Markov Models What are they? Why are they useful? The maths part Probability calculations Training optimising parameters Viterbi unseen sequences Real

More information

Machine Learning & Data Mining Caltech CS/CNS/EE 155 Hidden Markov Models Last Updated: Feb 7th, 2017

Machine Learning & Data Mining Caltech CS/CNS/EE 155 Hidden Markov Models Last Updated: Feb 7th, 2017 1 Introduction Let x = (x 1,..., x M ) denote a sequence (e.g. a sequence of words), and let y = (y 1,..., y M ) denote a corresponding hidden sequence that we believe explains or influences x somehow

More information

Dept. of Linguistics, Indiana University Fall 2009

Dept. of Linguistics, Indiana University Fall 2009 1 / 14 Markov L645 Dept. of Linguistics, Indiana University Fall 2009 2 / 14 Markov (1) (review) Markov A Markov Model consists of: a finite set of statesω={s 1,...,s n }; an signal alphabetσ={σ 1,...,σ

More information

Recall: Modeling Time Series. CSE 586, Spring 2015 Computer Vision II. Hidden Markov Model and Kalman Filter. Modeling Time Series

Recall: Modeling Time Series. CSE 586, Spring 2015 Computer Vision II. Hidden Markov Model and Kalman Filter. Modeling Time Series Recall: Modeling Time Series CSE 586, Spring 2015 Computer Vision II Hidden Markov Model and Kalman Filter State-Space Model: You have a Markov chain of latent (unobserved) states Each state generates

More information

A Higher-Order Interactive Hidden Markov Model and Its Applications Wai-Ki Ching Department of Mathematics The University of Hong Kong

A Higher-Order Interactive Hidden Markov Model and Its Applications Wai-Ki Ching Department of Mathematics The University of Hong Kong A Higher-Order Interactive Hidden Markov Model and Its Applications Wai-Ki Ching Department of Mathematics The University of Hong Kong Abstract: In this talk, a higher-order Interactive Hidden Markov Model

More information

STATS 306B: Unsupervised Learning Spring Lecture 5 April 14

STATS 306B: Unsupervised Learning Spring Lecture 5 April 14 STATS 306B: Unsupervised Learning Spring 2014 Lecture 5 April 14 Lecturer: Lester Mackey Scribe: Brian Do and Robin Jia 5.1 Discrete Hidden Markov Models 5.1.1 Recap In the last lecture, we introduced

More information

Hidden Markov models

Hidden Markov models Hidden Markov models Charles Elkan November 26, 2012 Important: These lecture notes are based on notes written by Lawrence Saul. Also, these typeset notes lack illustrations. See the classroom lectures

More information

Computational Genomics and Molecular Biology, Fall

Computational Genomics and Molecular Biology, Fall Computational Genomics and Molecular Biology, Fall 2011 1 HMM Lecture Notes Dannie Durand and Rose Hoberman October 11th 1 Hidden Markov Models In the last few lectures, we have focussed on three problems

More information

Dynamic Approaches: The Hidden Markov Model

Dynamic Approaches: The Hidden Markov Model Dynamic Approaches: The Hidden Markov Model Davide Bacciu Dipartimento di Informatica Università di Pisa bacciu@di.unipi.it Machine Learning: Neural Networks and Advanced Models (AA2) Inference as Message

More information

Introduction to Machine Learning. Maximum Likelihood and Bayesian Inference. Lecturers: Eran Halperin, Yishay Mansour, Lior Wolf

Introduction to Machine Learning. Maximum Likelihood and Bayesian Inference. Lecturers: Eran Halperin, Yishay Mansour, Lior Wolf 1 Introduction to Machine Learning Maximum Likelihood and Bayesian Inference Lecturers: Eran Halperin, Yishay Mansour, Lior Wolf 2013-14 We know that X ~ B(n,p), but we do not know p. We get a random sample

More information

Clustering K-means. Clustering images. Machine Learning CSE546 Carlos Guestrin University of Washington. November 4, 2014.

Clustering K-means. Clustering images. Machine Learning CSE546 Carlos Guestrin University of Washington. November 4, 2014. Clustering K-means Machine Learning CSE546 Carlos Guestrin University of Washington November 4, 2014 1 Clustering images Set of Images [Goldberger et al.] 2 1 K-means Randomly initialize k centers µ (0)

More information

Gaussian Mixture Models, Expectation Maximization

Gaussian Mixture Models, Expectation Maximization Gaussian Mixture Models, Expectation Maximization Instructor: Jessica Wu Harvey Mudd College The instructor gratefully acknowledges Andrew Ng (Stanford), Andrew Moore (CMU), Eric Eaton (UPenn), David Kauchak

More information

University of Cambridge Engineering Part IIB Module 4F10: Statistical Pattern Processing Handout 2: Multivariate Gaussians

University of Cambridge Engineering Part IIB Module 4F10: Statistical Pattern Processing Handout 2: Multivariate Gaussians Engineering Part IIB: Module F Statistical Pattern Processing University of Cambridge Engineering Part IIB Module F: Statistical Pattern Processing Handout : Multivariate Gaussians. Generative Model Decision

More information

Lecture 7 Sequence analysis. Hidden Markov Models

Lecture 7 Sequence analysis. Hidden Markov Models Lecture 7 Sequence analysis. Hidden Markov Models Nicolas Lartillot may 2012 Nicolas Lartillot (Universite de Montréal) BIN6009 may 2012 1 / 60 1 Motivation 2 Examples of Hidden Markov models 3 Hidden

More information