edu.cmu.sphinx.linguist.acoustic.tiedstate
Class Sphinx3Loader

java.lang.Object
  edu.cmu.sphinx.linguist.acoustic.tiedstate.Sphinx3Loader

All Implemented Interfaces:

Loader, Configurable

public class Sphinx3Loader
extends java.lang.Object
implements Loader

Loads a tied-state acoustic model generated by the Sphinx-3 trainer.

The acoustic model should be packaged in a JAR file or should be unpacked on disk. The dictionary and language model files are not required to be in the package. You can specify their locations separately. Properties of the model should be either configure in a config file or stored with the model in "model.props" text file. The "model.props" has the priority over the config file. It is a file of key-value pairs, loadable as a Java Properties file. It should minimally contain the following properties:

• dataLocation - this specifies the directory where the actual model data files are, relative to the model implementation class
• modelDefinition - this specifies the location where the model definition file is, relative to the model implementation class

The actual model data files are named "means", "variances", "transition_matrices", "mixture_weights" for binary versions, or prepended with ".ascii" for the ASCII versions.

As an example, lets look at the Wall Street Journal acoustic model JAR file, which is located at the sphinx4/lib directory. If you run "jar tvf lib/WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz.jar", you will find that its internal structure looks roughly like:

 WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz.jar
   |
   +- edu
       |
       +- cmu
           |
           +- sphinx
               |
               +- model
                   |
                   + acoustic
                      |
                      +- model.props
                      |
                      +- WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz.class
                      |
                      +- WSJLoader.class
                      |
                      +- cd_continuous_8gau
                      |   |
                      |   +- means
                      |   +- variances
                      |   +- mixture_weights
                      |   +- transition_matrices
                      |
                      +- dict
                      |   |
                      |   +- alpha.dict
                      |   +- cmudict.0.6d
                      |   +- digits.dict
                      |   +- fillerdict
                      |
                      +- etc
                          |
                          +- WSJ_clean_13dCep_16k_40mel_130Hz_6800Hz.4000.mdef
                          +- WSJ_clean_13dCep_16k_40mel_130Hz_6800Hz.ci.mdef
                          +- variables.def
 

The model.props file looks like (note how the 'dataLocation' and 'modelDefinition' properties are defined relative to the WSJ_clean_13dCep_16k_40mel_130Hz_6800Hz.class):

 description = Wall Street Journal acoustic models
 isBinary = true
 featureType = cepstra_delta_doubledelta
 vectorLength = 39
 sparseForm = false
 <p/>
 numberFftPoints = 512
 filters = 40
 gaussians = 8
 maxFreq = 6800
 minFreq. = 130
 sampleRate = 16000
 

 dataLocation = cd_continuous_8gau
 modelDefinition = etc/WSJ_clean_13dCep_16k_40mel_130Hz_6800Hz.4000.mdef
 

Note that although most of the properties of this class are already defined in the model.props file, it is still possible (but not recommended) to override those values by specifying them in the configuration file.

Field Summary

static java.lang.String	MODEL_VERSION Supports this version of the acoustic model
static java.lang.String	PROP_DATA_LOCATION Subfolder where the acoustic model can be found
static java.lang.String	PROP_IS_BINARY Specifies whether the model to be loaded is in ASCII or binary format
static java.lang.String	PROP_LOG_MATH The log math component for the system.
static java.lang.String	PROP_MC_FLOOR Mixture component score floor.
static java.lang.String	PROP_MODEL The name of the model definition file (contains the HMM data)
static java.lang.String	PROP_MW_FLOOR Mixture weight floor.
static java.lang.String	PROP_PROPERTIES_FILE The SphinxProperty for the name of the acoustic properties file.
static java.lang.String	PROP_SPARSE_FORM The SphinxProperty specifying whether the transition matrices of the acoustic model is in sparse form, i.e., omitting the zeros of the non-transitioning states.
static java.lang.String	PROP_UNIT_MANAGER The unit manager
static java.lang.String	PROP_USE_CD_UNITS The SphinxProperty specifying whether context-dependent units should be used.
static java.lang.String	PROP_VARIANCE_FLOOR Variance floor.
static java.lang.String	PROP_VECTOR_LENGTH The SphinxProperty for the length of feature vectors.

Constructor Summary

Sphinx3Loader()

Sphinx3Loader(java.lang.String propsFile, LogMath logMath, UnitManager unitManager, boolean isBinary, boolean sparseForm, int vectorLength, java.lang.String model, java.lang.String dataDir, float distFloor, float mixtureWeightFloor, float varianceFloor, boolean useCDUnits)

Method Summary

java.util.Map<java.lang.String,Unit>	getContextIndependentUnits() Returns the map of context indepent units.
HMMManager	getHMMManager() Returns the HMM Manager for this loader.
int	getLeftContextSize() Returns the size of the left context for context dependent units.
Pool<float[]>	getMeansPool() Gets the pool of means for this loader.
Pool<float[][]>	getMeansTransformationMatrixPool() Gets the means transformation matrix pool.
Pool<float[]>	getMeansTransformationVectorPool() Gets the means transformation vectors pool.
Pool<float[]>	getMixtureWeightPool() Gets the mixture weight pool.
int	getRightContextSize() Returns the size of the right context for context dependent units.
Pool<Senone>	getSenonePool() Gets the senone pool for this loader.
float[][]	getTransformMatrix() Gets the transformation matrix.
Pool<float[][]>	getTransitionMatrixPool() Gets the transition matrix pool.
Pool<float[]>	getVariancePool() Gets the variance pool.
Pool<float[][]>	getVarianceTransformationMatrixPool() Gets the variance transformation matrix pool.
Pool<float[]>	getVarianceTransformationVectorPool() Gets the variance transformation vectors pool.
void	load() Loads the acoustic model.
void	logInfo() logs information about this loader
void	newProperties(PropertySheet ps) This method is called when this configurable component needs to be reconfigured.

Methods inherited from class java.lang.Object

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

Field Detail

PROP_LOG_MATH

@S4Component(type=LogMath.class)
public static final java.lang.String PROP_LOG_MATH

The log math component for the system.

See Also:

Constant Field Values

PROP_UNIT_MANAGER

@S4Component(type=UnitManager.class)
public static final java.lang.String PROP_UNIT_MANAGER

The unit manager

See Also:

Constant Field Values

PROP_IS_BINARY

@S4Boolean(defaultValue=true)
public static final java.lang.String PROP_IS_BINARY

Specifies whether the model to be loaded is in ASCII or binary format

See Also:

Constant Field Values

PROP_MODEL

@S4String(mandatory=false,
          defaultValue="mdef")
public static final java.lang.String PROP_MODEL

The name of the model definition file (contains the HMM data)

See Also:

Constant Field Values

PROP_DATA_LOCATION

@S4String(mandatory=false,
          defaultValue="data")
public static final java.lang.String PROP_DATA_LOCATION

Subfolder where the acoustic model can be found

See Also:

Constant Field Values

PROP_PROPERTIES_FILE

@S4String(defaultValue="model.props")
public static final java.lang.String PROP_PROPERTIES_FILE

The SphinxProperty for the name of the acoustic properties file.

See Also:

Constant Field Values

PROP_VECTOR_LENGTH

@S4Integer(defaultValue=39)
public static final java.lang.String PROP_VECTOR_LENGTH

The SphinxProperty for the length of feature vectors.

See Also:

Constant Field Values

PROP_SPARSE_FORM

@S4Boolean(defaultValue=true)
public static final java.lang.String PROP_SPARSE_FORM

The SphinxProperty specifying whether the transition matrices of the acoustic model is in sparse form, i.e., omitting the zeros of the non-transitioning states.

See Also:

Constant Field Values

PROP_USE_CD_UNITS

@S4Boolean(defaultValue=true)
public static final java.lang.String PROP_USE_CD_UNITS

The SphinxProperty specifying whether context-dependent units should be used.

See Also:

Constant Field Values

PROP_MC_FLOOR

@S4Double(defaultValue=0.0)
public static final java.lang.String PROP_MC_FLOOR

Mixture component score floor.

See Also:

Constant Field Values

PROP_VARIANCE_FLOOR

@S4Double(defaultValue=9.999999747378752E-5)
public static final java.lang.String PROP_VARIANCE_FLOOR

Variance floor.

See Also:

Constant Field Values

PROP_MW_FLOOR

@S4Double(defaultValue=1.0000000116860974E-7)
public static final java.lang.String PROP_MW_FLOOR

Mixture weight floor.

See Also:

Constant Field Values

MODEL_VERSION

public static final java.lang.String MODEL_VERSION

Supports this version of the acoustic model

See Also:

Constant Field Values

Constructor Detail

Sphinx3Loader

public Sphinx3Loader(java.lang.String propsFile,
                     LogMath logMath,
                     UnitManager unitManager,
                     boolean isBinary,
                     boolean sparseForm,
                     int vectorLength,
                     java.lang.String model,
                     java.lang.String dataDir,
                     float distFloor,
                     float mixtureWeightFloor,
                     float varianceFloor,
                     boolean useCDUnits)

Sphinx3Loader

public Sphinx3Loader()

Method Detail

newProperties

public void newProperties(PropertySheet ps)
                   throws PropertyException

Description copied from interface: Configurable

This method is called when this configurable component needs to be reconfigured.

Specified by:

newProperties in interface Configurable

Parameters:

ps - a property sheet holding the new data

Throws:

PropertyException - if there is a problem with the properties.

load

public void load()
          throws java.io.IOException

Description copied from interface: Loader

Loads the acoustic model.

Specified by:

load in interface Loader

Throws:

java.io.IOException - if an error occurs while loading the model

getContextIndependentUnits

public java.util.Map<java.lang.String,Unit> getContextIndependentUnits()

Description copied from interface: Loader

Returns the map of context indepent units. The map can be accessed by unit name.

Specified by:

getContextIndependentUnits in interface Loader

Returns:

the map of context independent units

getMeansPool

public Pool<float[]> getMeansPool()

Description copied from interface: Loader

Gets the pool of means for this loader.

Specified by:

getMeansPool in interface Loader

Returns:

the pool

getMeansTransformationMatrixPool

public Pool<float[][]> getMeansTransformationMatrixPool()

Description copied from interface: Loader

Gets the means transformation matrix pool.

Specified by:

getMeansTransformationMatrixPool in interface Loader

Returns:

the pool

getMeansTransformationVectorPool

public Pool<float[]> getMeansTransformationVectorPool()

Description copied from interface: Loader

Gets the means transformation vectors pool.

Specified by:

getMeansTransformationVectorPool in interface Loader

Returns:

the pool

getVariancePool

public Pool<float[]> getVariancePool()

Description copied from interface: Loader

Gets the variance pool.

Specified by:

getVariancePool in interface Loader

Returns:

the pool

getVarianceTransformationMatrixPool

public Pool<float[][]> getVarianceTransformationMatrixPool()

Description copied from interface: Loader

Gets the variance transformation matrix pool.

Specified by:

getVarianceTransformationMatrixPool in interface Loader

Returns:

the pool

getVarianceTransformationVectorPool

public Pool<float[]> getVarianceTransformationVectorPool()

Description copied from interface: Loader

Gets the variance transformation vectors pool.

Specified by:

getVarianceTransformationVectorPool in interface Loader

Returns:

the pool

getMixtureWeightPool

public Pool<float[]> getMixtureWeightPool()

Description copied from interface: Loader

Gets the mixture weight pool.

Specified by:

getMixtureWeightPool in interface Loader

Returns:

the pool

getTransitionMatrixPool

public Pool<float[][]> getTransitionMatrixPool()

Description copied from interface: Loader

Gets the transition matrix pool.

Specified by:

getTransitionMatrixPool in interface Loader

Returns:

the pool

getTransformMatrix

public float[][] getTransformMatrix()

Description copied from interface: Loader

Gets the transformation matrix.

Specified by:

getTransformMatrix in interface Loader

Returns:

the matrix

getSenonePool

public Pool<Senone> getSenonePool()

Description copied from interface: Loader

Gets the senone pool for this loader.

Specified by:

getSenonePool in interface Loader

Returns:

the pool

getLeftContextSize

public int getLeftContextSize()

Description copied from interface: Loader

Returns the size of the left context for context dependent units.

Specified by:

getLeftContextSize in interface Loader

Returns:

the left context size

getRightContextSize

public int getRightContextSize()

Description copied from interface: Loader

Returns the size of the right context for context dependent units.

Specified by:

getRightContextSize in interface Loader

Returns:

the left context size

getHMMManager

public HMMManager getHMMManager()

Description copied from interface: Loader

Returns the HMM Manager for this loader.

Specified by:

getHMMManager in interface Loader

Returns:

the HMM Manager

logInfo

public void logInfo()

Description copied from interface: Loader

logs information about this loader

Specified by:

logInfo in interface Loader