eq4 / hmmduration

Python Hidden Markov Models framework. Adapted for computationally optimal Viterbi forced alignment. Added Explicit Duration model

HMM

A numpy/python-only Hidden Markov Models framework. No other dependencies are required.

This implementation (like many others) is based on the paper: "A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition, LR RABINER 1989"

Major supported features:

• Discrete HMMs
• Continuous HMMs - Gaussian Mixtures
• Supports a variable number of features
• Easily extendable with other types of probablistic models (simply override the PDF. Refer to 'GMHMM.py' for more information)
• Non-linear weighing functions - can be useful when working with a time-series

Open concerns:

• Examples are somewhat out-dated
• Convergence isn't guaranteed when using certain weighing functions

---

Duration-model extension : done by [email protected]

• fixed underflow due to multiplication of probabilities ( handled by sum log(probs) )

• posteriors of GMM mixtures are computed using scikit learn

• added oracle test: Oracle test allow to check if model is able to perform perfect alignment on ground truth timestamps of phonemes (e.g. replace observation posteriors with 1-s from ground truth) (described more here: http://www.terasoft.com.tw/conf/ismir2014/proceedings/T050_126_Paper.pdf ) hmm.examples.tests.test_oracle() - needs phoneme-level annotation

Algorithm parameters can be changed from class hmm.ParametersAlgo

TODO: use _ContinuousHMM.usePersistentProbs to store persistently .durationsMap to save time Now computed each time