Explicite Duration Modeling - Tcl

Duration Methods

• <object> puts

  List all the probabilities for every duration-range bucket. The listed probabilites should add up to one.

DurationSet Methods

• <object> puts

  List all the names of the single Duration objects that are subobject of the DurationSet object.

• <object> read <filename>

  Read duration model names from the named file. The duration models in the file are added to the those that are already in the given object. The 'old' duration models are not removed.

• <object> write <filename>

  Expects one argument, the name of the file into which it should write the names of the duration models in the given object.

• <object> add <name> <probs> [-count <count>]

  Expects the name of a duration model which will be added to the ones that are already in the DurationSet object. The <probs> argument is a list of float numbers, one number for each bucket of the DurationSet. The numbers should add up to one. If they don't they will be normalized, and if their sum is too far away from 1.0 a warning is issued. The optional <count> argument specifies the number of times the duration model was seen during training. This value is used for merging and clustering distribution models.

• <object> delete <name*>

  Expects the name of a duration model which will be removed from the DurationSet object. Please keep in mind, that deleting a duration model whose index is n does change the indices of the models with index m>n. Adding and deleting duration models should be finished before a DurationSet object is given to any other object for indexing.

• <object> index <name*>

  Expects any number of duration model names. The method will return the list of indices of the named models.

• <object> name <index*>

  Expects any number of integers and returns the names of the duration models whose index they represent. If an index is less than zero or greater than the greatest index in the object the resulting name will be (null). Please keep in mind, that deleting subobjects from an object can change the indices of other subobjects.

• <object> prob <name> <frameN>

  Print the probability that the duration of the given model is <frameN>.

• <object> scale <factor>

  Define a multiplicative scaling factor to the log-probs that are stored in the given object. I.e., after a scaling factor <factor> has been applied the returned scores will not be
  log P(duration=FrameN)
  but
  <factor> * log P(duration=FrameN).
  Subsequently applied scalings multiply and don't replace the current scaling.

• <object> prob <name1> <name2>

  Print the distance (increase of entropy if the two models were merged) of the two given models. The distance is computed correctly, taking into account that different buckets can have different sizes.

• <object> accu <path> <hmm> [-factor <factor>]

  Accumulate the training data specified by the given path and HMM object. Use the optionally given training factor.

• <object> update

  Adjust the duration models parameters according to the current accumulators. The accumulators are not emptied after this operation. Duration models whose accumulators did not get enough counts will not be updated, the updating can use floor values and a momentum (as is the same with distributions). See the configurable parameters below for details on these.

• <object> createAccus

  Allocate an accumulator structure for every duration model in the object.

• <object> freeAccus

  Deallocate all accumulator structures in the object.

• <object> clearAccus

  Reset all accumulators of the object. This is done automatically after the accumulators have been allocated. It is not done after the duration parameters have been updated.

• <object> loadAccus <filename>

  Load (and add to the current values) duration model accumulators from the named file.

• <object> saveAccus <filename>

  Save the current accumulator values into the named file.

• <object> putsAccu [<name>]

  Print out the current values of all (or just of the optionally named model's) accumulators.

Configurable Parameters for the Duration Object

• -count <integer>
  is the occurrence count of the model (used for merging models or computing their distrances.

• -valN <integer> (read only)
  specifies the number of buckets used.

• -name <string> (read only)
  is the name of the model.

Configurable Parameters for the DurationSet Object

• -useN <integer>
  is the count of super-objects that are using the cofigured object.

• -itemN <integer>
  is the number of items in the configured object, i.e. the number of Duration objects in a DurationSet object.

• -blkSize <integer>
  is the blocksize of the used generic list.

• -commentChar <character>
  If (while reading a file) this character is the first character of a line, then that line will be ignored. When writing a file this character will be used to initiate a comment header.

• -floor <float>
  is the minimum count for each bucket. If the accumulated count in one bucket is below this floor value, it will be set to the floor value, prior to updating the parameters.

• -minCount <float>
  is the minimum count for each bucket that is desired for an update. If the accumulated count in one bucket is below this value, the corresponding parameters will not be updated, they will remain unchanged.

• -momentum <float>
  When the parameters of the Duration object are updated, then they will be replaced by the current value multiplied with the momentum plut the value computed from the accumulators times one minus the momentum.