Visual HMM

Hidden Markov Model (HMM) is a generative model technique used in speech recognition, computational biology and many other research fields and industry. Visual HMM offers a way to visualize HMM parameters in feature space. Currently, with Quick HMM Look tool, a researcher may open HMM definitions and analyze visual presentation of emitting states for each model like a phoneme for speech recognition system.

System requirements

The Quick HMM Look tool is developed for MS Windows. Wine is required to run it on Linux or Macintosh. We plan to support more platforms natively.

Download

Quick HMM Look for Windows
Visual HMM sources
HMM samples for Ukrainian speech recognition build on single and multiple speaker data bases. Model names are context-independent phonemes containing both stressed (marked with "1" like "a1") and unstressed vowels.

HMM visualization tool Quick HMM Look (Wine on Ubuntu)

Features

Loading HMMs written in HTK text format.
Switch to models selected from the droplist.
Emitting states are shown left-to-right. Only GMMs with diagonal covariance matrix are supported within states.
Each state consists of three top-down parts corresponding to basic coefficents like MFCC, LPC or PLP, velocity (delta) and acceleration (delta-delta) dynamic coefficients.
Basic coefficients, basic energy coefficient, velocity and acceleration have specific scale that is same throughout all models. Scale limits are denoted (in parentheses for basic energy coefficient).
GMM's means are shown as bars that might be switched on and off.
GMM's variances are presented in gradient form.
Stretch is helpful to see mean locations in details (available from View menu). Meanwhile, this function is useful now mostly for basic coeficients.

Stretched basic model coefficients on Quick HMM Look

Comments, advices

Feature space is suggested to contain delta and delta-delta coefficients accomplished with energy.
Only HMMs in HTK text format can be opened.
Inverted colors are more appropriate for a printable figure (Edit->Invert colors).
Do not try to find a horizontal zoom: it is not implemented yet.
Please ignore Edit->Cut/Copy/Paste items in menu. Currently, they do nothing. Same situation is for New, Save and Print commands.

Plans

We hope to develop the project continuously. The following features will be available:

Model comparizon.
GMM visual evolution.
Connecting coefficients for the same GMM.
State transition presentation.
Mapping selected dimensions of the feature space to the plane.
Speech signal trajectory visualization on the mapped features space.
Improved user interface.

Most features above will be implemented in an open-source cross-platform HMM visualization module.

Keep an eye on the forum topic Speech Recognition Theoretical Basis to learn news and make suggestions.

Credits

The idea of HMM visualization is contributed by Prof. Taras Vintsiuk in 2005.