6. The cost250 Tcl package

6.1 Using the package

The example program assumes that the experiment definition files are organized according to the following convention. A set of three experiment definition files are identified by a tag and are supposed to be stored in a directory ../experiment/tag. The three files should have names

• os_wld.exp - for training of a world model (or any set of non-client models used by the recognition engine).
• es.exp - for enrollment
• ts.exp - for verification attempts

The commands are described below from an object-oriented point-of-view. A description of the object-oriented programming style we have tried to follow with Tcl can be found here.

This command will perform all operations listed in the file experimentFileName. The given database and engine object will be used to implement the operations. Enrolled models are saved in directory modelDir and results are saved in outfile. Default value for modelDir is the current directory and the default value for outfile is results.llk in the current directory.

The required format for the experiment definition file is described in section File Format Description.

The source code for the Experiment class can be found in tcl/runexp.tcl in the source distribution.

where datadir is the base directory, fileTag uniquely identifies a single speech file, and alwExt is the filename extension. Polycost satisfies this condition if all subject directories are copied to disk into the same directory. It is not satisfied with speech data being stored on two separate CR-ROMs. Furthermore, the Reference System recognition engine currently assumes that the speech file supplied by the database object is a headerless, 8 kHz A-law file. The current task of the database class is to map a file tag onto a complete speech filename. This should be changed in future versions of the Reference System to enable the use of other speech file formats with other databases than Polycost; the database object should return the actual, decoded, speech data instead of the name of a file.

A database object is created with the command

It returns the name of the created object, databaseName, defined in the global namespace. The name will be something like "::db01". Multiple database objects can be created and they will all have unique names.

This is the typical life cycle of a database object:

1. The object is created by the Database::new command.
2. Default property values are optionally overridden with the set method.
3. The init method is called after all properties have been set.
4. The database object is passed as an argument to one or more Experiment::run commands and subsequently to a recognition engine object. The engine object will call the database object's filename method to map a speech file tag to a filename.
5. The destroy method is called when the object is no longer needed.

databaseName set propertyName value

Sets the value of a property. A property always has a default value which can be overridden by calling this method. Table 1 lists the properties that are defined for a database object and their default values.

Initializes the object. This method should always be called after set has been called and before any other methods are called for this object.

Prints the property values of the object. An optional output channel can be given. Default output channel is stdout.

Returns the name of the database root directory.

Returns the full file name for the speech file indicated by fileTag.

Destroys the object. This method should always be called when the object is no longer needed.

property	default	description
alwExt	.alw	the filename extension for speech files in the database
trace	0	currently not used.

Table 1. Properties in a Database object.

• The speech signal is pre-emphasized using a factor of 0.97 and divided into 100 overlapping frames per second. The analysis window for each frame is 25 ms and, hence, we have a 60% overlap between consecutive frames. For each frame a 12th order LPC-cepstrum vector is computed.
• A 12-dimensional VQ codebook with 64 codewords constitutes a client speaker model. Another VQ codebook, also with 64 codewords, constitutes a non-client world model and is used for score normalization. Codebooks are trained using the LBG-algorithm.
• The output from a verification attempt is a score value computed as the non-client VQ distortion divided by the client VQ distortion.

where databaseName is the name of a database object from which speech files can be retrieved. The command returns the name of the created object, engineName, defined in the global namespace. The name will be something like "::rec01". Multiple engine objects can be created and they will all have unique names.

This is the typical life cycle of an engine object:

1. The object is created by the ReferenceSystem::new command.
2. Default property values are optionally overridden with the set method.
3. The init method is called after all properties have been set.
4. The engine object is passed as an argument to the Experiment::run command. This command will call the engine object's enrollment and verification methods to perform a list of operations.
5. The destroy method is called when the object is no longer needed.

engineName set propertyName value

Sets the value of a property. A property always has a default value which can be overridden by calling this method. Table 2 lists the properties that are defined for an engine object and their default values.

Initializes the engine. This method should always be called after set has been called and before any other methods are called for this object.

Prints the property values of the object. An optional output channel can be given. Default output channel is stdout.

Performs one enrollment operation, where a model is trained for the given identity based on the speech files indicated by fileTags (a Tcl list). The produced model is saved in directory outDir.

Performs one verification trial, where a speaker claims to have a given identity using the speech files indicated by fileTags (a Tcl list) to support the claim. The actual identity of the speaker is speaker. The result is printed to resultChannel.

Destroys the object. This method should always be called when the object is no longer needed.

property*	default	description
lpcConfig*	-f200 -o60 -p12 -a0.97 -s1	configuration options to lin2parProgram. The available configuration options are described in Table 3.
cbSize*	64	size of codebook to create during an enrollment operation
normalization*	1	1 = use score normalization; 0 = no score normalization
nonClientModels*	W	a list of non-client model names separated with space (for example: "W", or "F M")
binDir	""	where executable VQST files reside. May be empty ("") if they are in the current search path.
alw2linProgram*	alw2lin	binary program that decodes A-law sample to linear scale
lin2parProgram*	lpccep	binary program that parameterizes a speech file
concatProgram*	concat	binary program that concatenates parameterized files
gencbProgram*	gencb	binary program that trains a speaker model (codebook)
vqtestProgram*	vqtest	binary program that computes a score for a test utterance against a speaker model
tmpDir	/tmp	where to create temporary files
cliDir	cli	where to find client model files (codebooks)
refDir	ref	where to find non-client model files (codebooks)
linearExt	.lin	filename extension for sample files with linear scale
paramExt	.lpc	filename extension for parameter files
modelExt	.cb	filename extension for client and non-client model files (codebooks)
uniqTag		set automatically by the new command
lpcConfigFileName		set automatically by the new command
trace	0	Set to 1 for trace output; 0 for no trace output.

Table 2. Properties in a ReferenceSystem object. ^*These properties must have their default values for the system to perform as a reference system in the sense that it produces the same results in all sites.

option	default	description
-f	256	the speech frame size in samples
-o	50[%]	the overlap between the neighbouring speech frames in percentage
-p	12	the LPC-cepstrum vector size
-t	0.01	the absolute energy threshold: this parameter is used to discard any speech frame whose absolute average energy is less than the given value
-a	0.95	the pre-emphasis coefficient
-h	0	header size (in number of integers) : its default value is "0" meaning use all the speech samples in the input file. If it is set to any non-zero value, that many samples at the top of the speech file will be ignored. Note that when the alw2lin program is used to produce the input to the LPC-cepstrum program, header size must be 0.
-s	0	a trace information parameter: if set to zero, info on various parameter values will be printed. If set to "1" no parameter values are printed (silent mode).

Table 3. Configuration options for the LPC-cepstrum program.