This is a Python script that enables visualization opeartions with the models.
Parses into a python class the HMM monophone models from htk. 


Operations:
 - comparison of two different models for the same phoneme. by calculating euclidean distance between means of model. 
@ file compare.compareModels.py
 - visualize features: plots feature vectors can plot two features in the same plot. 
@ file compare.visualize.py

The script was developed originally by W.J. Maaskant
(w.j.maaskant(at)student.utwente.nl) for Fraunhofer IAIS, Germany to perform
a comparison of HTK, ISIP Prototype, Julius and HTK. 

EDITED by Georgi Dzhambazov 
georgi.dzhambazov@upf.edu 


See the LICENSE file for the licensing terms.


Requirements
------------
The following Python modules are required:

- PLY
try: pip install ply

This script was developed and tested on openSUSE 10.2.

Capabilities and Limitations
----------------------------
This section contains information about the capabilities and limitations of the
converter. Some limitations are caused by the converter itself, some by
limitations of Sphinx 3.

- It is assumed that the HTK model file is correct. No checks are made on
  whether all vectors have the same size, for example.

- The following macro types are supported: ~h ~o ~s ~t ~u ~v. The content of
  ~o is ignored.

- Feature vectors must have 39 dimensions.

- Only 1 feature stream is supported.

- All models must have the following structure, visualized by a transition
  matrix:
   | 1 2 3 4 5
  -+----------
  1| 0 1 0 0 0
  2| 0 * * * *
  3| 0 * * * *
  4| 0 * * * *
  5| 0 0 0 0 0

  ('*' stands for a probability >= 0.)
  
  The number of states may be varied.

  The first and last state of the models must be non-emitting.

- All states must have the same number of Gaussian mixtures.

- A monophone model must exist for the base phone of every triphone model.
  Monophone models with no triphones with that monophone as base phone are
  allowed.

- Biphone models are not supported.

- If more than 1 mixture is used, state definitions must include the
  <NUMMIXES> tag. If only 1 mixture is used, state definitions MUST NOT
  include the <NUMMIXES> tag.

- All vectors (mean, variance, transition probabilities) must be written as
  a floating point value in this (regular expression) format:
  -?\d.\d\d\d\d\d\de[+-]\d\d

- Models with base phone "sil" are assumed to be fillers.

Some limitations arise from limitations in Sphinx compared to HTK, and
(most) limitations arise from the fact that the script was created to
accomodate my models. However, it's not difficult to change the rules and
support more features of the HTK model file format.

What you have to do:

- Add or remove the filler attribute to models in the model definitions
  file, where necessary.