Acoustic Modeling

onyx.am – Package providing Gaussian-based models

Modeling is done with three classes:

• DummyModel
• GaussianMixtureModel
• Hmm

The first three are all derived from GaussianModelBase.

In addition, there are several other classes supporting operations on collections of models:

• GmmMgr - a collection of GaussianMixtureModels or DummyModels
• HmmMgr - a collection of Hmms
• UttLattice - a structure for doing embeddded training of Hmms
• SimpleClassifier - a collection of 1-component GaussianMixtureModels with associated labels
• AdaptingGmmClassifier - a collection of GaussianMixtureModels with associated labels
• HmmClassifier - a collection of Hmms with associated labels
• AdaptingGmmClassProcessor - a processor wrapper around AdaptingGmmClassifier

Scoring in the Acoustic Modeling package

Say something about scoring here.

Adaptation in the Acoustic Modeling package

Adaptation interfaces are available at many levels in the AM package. Also, there are several styles of interface, reflecting both the nature of the underlying algorithms and user demand.

Class	One-shot?	Begin/End?	Relevance	Algorithm	Notes
SimpleGaussianModel	Yes	Yes	Yes	Closed form	The simplest case
GaussianMixtureModel	Yes	No	Yes	EM
Hmm	Yes	Yes	No	Baum-Welch	See also UttLattice
GmmMgr	No	Yes	No	EM
HmmMgr	No	Partial	No	??
UttLattice	Yes	??	No	Baum-Welch
AdaptingGmmClassifier	Yes	No	Yes	EM
AdaptingGmmClassProcessor	Yes	No	Yes	EM

Notes:

1. The GaussianMixtureModel adapt() function can do multiple EM iterations over a single collection of datapoints. Each iteration re-estimates the weights, means, and variances used in the next iteration.
2. The GmmMgr class supports accumulation and the application of accumulators. This is used by the Hmm and UttLattice classes in their implementation of Baum-Welch training of Hmms. It has the potential to be used for an external EM implementation, but currently there isn’t one.
3. Note about HmmMgr
4. Note about UttLattice
5. The AdaptingGmmClassifier class uses the one-shot adapt() function in GaussianMixtureModel to do adaptation.
6. The AdaptingGmmClassProcessor class uses the one-shot interface of the AdaptingGmmClassifier to do adaptation. An input event consists of a single label and a sequence of training frames; these frames are then use together in several EM iterations (see note 1.)

Acoustic Modeling Modules

• onyx.am.gaussian – Simple Gaussian models and mixtures
• onyx.am.hmm – Hidden Markov Models
• onyx.am.bwtrainer – Baum-Welch trainer for HMMs - see class UttLattice
• onyx.am.modelmgr – Model managers at various levels.
• onyx.am.hmm_mgr – A manager for Hmms. Also read and write functions for acoustic models.
• onyx.am.classifier – Classifiers using GaussianMixtureModels
• onyx.am.acctrack – Processor-based tracking and printing of classification accuracy
• onyx.am.testhmm – Various test drivers for Hmm class
• onyx.am.testhmm2 – Various test drivers for Hmm class; these use the log domain
• onyx.am.testbwt – >>> True
• onyx.am.gmm_trainer – Training Gaussian mixture models.