Description of all Modules

• Allophone Model Indexing
  This model is keeping a list of all acoustically distinguishable allophones, i.e. a sequence of senone and transition model indices for every allophone. It asures that the same allophonic acoustic model is indexed with the same index. The object type AModel is defined here.

• Codebooks
  Codebooks are used to store the means and covariance matrices of the densities of a mixture. The codebooks module defines the objects CodebookSet, which is a set of Codebook objects. The CodebookAccu object is a training data accumulator. The Covar object is a covariance matrix.

• Data Base
  The data base module defines the objects DBase and DBaseIdx. A DBase object maintains a set of entries, each of which has a key and a value. These data bases are stored in the file system and can be read from there. A DBaseIdx object maintains the positions of each entry.

• Dictionary
  The dictionary module defines the objects Dictionary and Word. It maintains a list of words (orhtographic spelling and a seuqence of monophone and tag names. It does not (yet) store entire word-dependent HMMs.

• Distributions
  The distributions module defines the objects DistribSet and Distrib. It is used to store and maintain the mixture weights of mixture densities.

• Duration Modeling
  During testing every phone can be penalized by some value which depends on its duration. The duration modeling module can learn and maintain duration histograms. Duration models can be clustered in a tree just the same way as topologies or distributions.

• Error Handling
  This module is rather small. It manages the flow of error messages and warnings and info messages. The actual handling of error messages can be done in Tcl.

• Features
  The feature module defines the Feature object. Multiple features can be stored in one FeatureSet object. The feature module offers many operations that can be performed on features (like computing FFTs, matrix multiplication, cutting and pasting around in the feature space etc.).

• Generic List
  The generic list module offers functions for maintaing lists of anything. Many modules define objects type which make heavy use of generic lists.

• Generic Tree
  The generic tree is only used as a basis for senone and topology trees. The object type Tree is defined here.

• Hidden Markov Model
  The HMM Module defines the object type HMM, which contains the entire topology of an utterance (word-graph, phone-graph, state-graph).

• Hypotheses
  
  Using JANUS, you probably never need to create a HypoList object. However, the Search object contains a HypoList to store it's results. If you want to manipulate those, the methods used for a HypoList will be interesting for you.

• Interface C/Tcl
  The C/Tcl interface module does not define any object itself. This module offers many C-functions and tools for other modules to declare and maintain object types, as well as handling error messages. Click here if you want to know how to implement new modules and object classes.

• Labels of Speech Segments
  The labels module defines the Label and LabelDesc object. A LabelDesc object is a collection of many other subobjects (dictionary, senone tree, etc.) that are needed to make sure that the meaning of labels can be interpreted correctly. The Label object maintains a database of labels (using a DBaseIdx).

• LDA
  
  A LDA object can be used to calculate total and within class scatter matrices using a Path object and data from a FeatureSet. With this information a transformation matrix A can be obtained that can be used afterwards to reduce the dimension of the input feature.

• Language Model
  
  A Lm object contains a statistical language model. It's defined only over a Vocab object. Currently, only trigrams, bigrams and unigrams are implemented.

• Machine-Independent I/O
  
  This module implements functions that can be used for writing and reading binary data across different platforms, making sure that the data remain the same.

• Matrices
  
  Matrices are used all over the system. This module provides different kinds of matrices and vector objects and offers functions to work with them.

• Path (output of aligment)
  A Path object can be filled by a forced alignment procedure or the labels module. It contains alignment information (which only make sense if the corresponding HMM object is available).

• Phones
  The phones module defines the objects Phones and PhonesSet used for assigning unique indices to monophone names and to define qeustions about an allophone context.

• Polyphone Trees
  The polyphone trees (ptree) module defines the object Ptree used for managing arbitrary wide contexed polyphones.

• Rewriting of Names
  The rewriting module keeps simple rewriting rules, each of which is a pair of identifyers, defining which identifyer should be renamed by what. The codebooks and distributions module make use of this. Here the object types Rewrite and RewriteSet are defined.

• Sample Set
  For extracting sample vectors from the training data and sorting them into class-dependent files, you need some buffers. The accumulation and dumping of such samples is done in this module. Other Services include the computation of initial codebooks with k-means or "neural gas".

• Search (Decoding)
  The search module defines the Search object and other subobjects that are used for decoding (i.e. recognition).

• Senones
  The senones module defines the objects SenoneSet and Senone. A senone set is just a collection of many senones. Each senone is representing an atomic acoustic model. The senones module offers transparent functions for computing emission probabilities, accumulating training data, and updating the acoustic parameters.

• Senone Tree
  The senone tree module defines the STree object which itself is defined over a generic tree. A senone tree is used to cluster senones according to an allophones context.

• Tags
  The tags module defines the object Tags which is used for indexing phoneme-tags (like 'relative word position', 'stress level', 'semantic or syntactic boundary').

• Topologies
  The topology module defines the TopoSet object and the Topo object. A topology is the definition of HMM states and transitions, it also defines a node of a senone tree from where to start for finding the senone that should be used for modelling the state.

• Topology Tree
  The topology tree module defines the TTree object which itself is defined over a generic tree. A topology tree is used to cluster HMM-topologies according to an allophones context.

• Transition Models
  The transition models module defines the objects Tm and TmSet. It is used to store and maintain the topologies of single HMM states, i.e. keeping a list of all transitions out of the state together with their probabilities.

• Viterbi Alignment
  The Viterbi module does not define any objects. It only implements the Viterbi alignment algorithm, i.e. traversing a given HMM and storing the path in a Path

• Vocabulary
  
  A Vocab Object stores the Recognition Vocabulary, defining common indecees for search and language module. If used for search it must be defined over a dictionary for valid pronounciations.

• ...