ocvolume.hmm
Class markov

java.lang.Object
  |
  +--ocvolume.hmm.markov

public class markov

extends java.lang.Object

last updated on June 15, 2002
description: this class represents a left-to-right Hidden Markov Model and its essential methods for speech recognition. The collection of methods include Forward-Backward Algorithm, Baum-Welch Algorithm, Scaling, Viterbi, etc.
calls: none
called by: volume, train
input: sequence of integers
output: probability

Field Summary

protected double[][]	alpha forward variable alpha
protected double[][]	beta backward variable beta
protected int[]	currentSeq current observation sequence
protected int	delta number of states the model is allowed to jump
protected int	len_obSeq length of observation sequence
protected int	num_obSeq number of observation sequence
protected int	num_states number of state in the model example: number of urns
protected int	num_symbols number of observation symbols per state example: how many different colour balls there are
protected int[][]	obSeq discrete set of observation symbols example: sequence of colour of balls
protected double[][]	output discrete output probability example: probability of a specific output from a state
protected double[]	pi initial state distribution example: which state is the starting state
int[]	q best state sequence
protected double[]	scaleFactor Scale Coefficient
protected double[][]	transition state transition probability example: probability from one state to another state

Constructor Summary

markov(int num_states, int num_symbols)
class constructor - used to create a left-to-right model with multiple observation sequences for training
calls: none
called by: trainHMM

markov(java.lang.String filepath)
class constructor - used to create a model from a saved file
calls: none
called by: volume, trainHMM

Method Summary

protected double	computeAlpha() calculate forward variable alpha calls: none called by: markov
protected void	computeBeta() calculate backward variable beta for later use with Re-Estimation method calls: none called by: markov
double	getProbability(int[] testSeq) returns the probability calculated from the testing sequence calls: none called by: volume
void	save(java.lang.String filepath) save HMM model to file calls: none called by: trainHMM
void	setNumObSeq(int k) set the number of training sequences calls: none called by: trainHMM
void	setObSeq(int[] observationSeq) set observation sequence calls: none called by: trainHMM
void	setTrainSeq(int[][] trainSeq) set training sequences for re-estimation step calls: none called by: trainHMM
void	setTrainSeq(int k, int[] trainSeq) set a training sequence for re-estimation step calls: none called by: trainHMM
void	train() train the hmm model until no more improvement calls: none called by: trainHMM
double	viterbi(int[] testSeq) viterbi algorithm used to get best state sequence and probability calls: none called by: volume

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

len_obSeq

protected int len_obSeq

length of observation sequence

num_states

protected int num_states

number of state in the model example: number of urns

num_symbols

protected int num_symbols

number of observation symbols per state example: how many different colour balls there are

delta

protected final int delta

number of states the model is allowed to jump

See Also:

Constant Field Values

obSeq

protected int[][] obSeq

discrete set of observation symbols example: sequence of colour of balls

currentSeq

protected int[] currentSeq

current observation sequence

num_obSeq

protected int num_obSeq

number of observation sequence

transition

protected double[][] transition

state transition probability example: probability from one state to another state

output

protected double[][] output

discrete output probability example: probability of a specific output from a state

pi

protected double[] pi

initial state distribution example: which state is the starting state

alpha

protected double[][] alpha

forward variable alpha

beta

protected double[][] beta

backward variable beta

scaleFactor

protected double[] scaleFactor

Scale Coefficient

q

public int[] q

best state sequence

Constructor Detail

markov

public markov(java.lang.String filepath)

class constructor - used to create a model from a saved file
calls: none
called by: volume, trainHMM

Parameters:

filepath - path of the file to load

markov

public markov(int num_states,
              int num_symbols)

class constructor - used to create a left-to-right model with multiple observation sequences for training
calls: none
called by: trainHMM

Parameters:

num_states - number of states in the model

num_symbols - number of symbols per state

Method Detail

viterbi

public double viterbi(int[] testSeq)

viterbi algorithm used to get best state sequence and probability
calls: none
called by: volume

Parameters:

testSeq - test sequence

Returns:

probability

getProbability

public double getProbability(int[] testSeq)

returns the probability calculated from the testing sequence
calls: none
called by: volume

Parameters:

testSeq - testing sequence

Returns:

probability of observation sequence given the model

computeAlpha

protected double computeAlpha()

calculate forward variable alpha
calls: none
called by: markov

Returns:

probability

computeBeta

protected void computeBeta()

calculate backward variable beta for later use with Re-Estimation method
calls: none
called by: markov

setNumObSeq

public void setNumObSeq(int k)

set the number of training sequences
calls: none
called by: trainHMM

Parameters:

k - number of training sequences

setTrainSeq

public void setTrainSeq(int k,
                        int[] trainSeq)

set a training sequence for re-estimation step
calls: none
called by: trainHMM

Parameters:

k - index representing kth training sequence

trainSeq - training sequence

setTrainSeq

public void setTrainSeq(int[][] trainSeq)

set training sequences for re-estimation step
calls: none
called by: trainHMM

Parameters:

trainSeq - training sequences

train

public void train()

train the hmm model until no more improvement
calls: none
called by: trainHMM

setObSeq

public void setObSeq(int[] observationSeq)

set observation sequence
calls: none
called by: trainHMM

Parameters:

observationSeq - observation sequence

save

public void save(java.lang.String filepath)

save HMM model to file
calls: none
called by: trainHMM

Parameters:

filepath - file location