Lecture 6 - Context-dependent phone models
In this lecture we discussed how HMMs are used for acoustic modelling, as context-dependent phone models.
- Different possible levels to of modelling with an HMM - constructing word HMMs as a sequence of phone HMMs
- Influence of phonetic context on the acoustics of a phone - coarticulation
- Divide and conquer strategy - multiple models for a single phone depending on the left and right phonetic context
- Triphone models, and modelling infrequent triphones by sharing Gaussians, states or models
- State clustering - states which are responsible for acoustically similar data are shared
- Choosing which contexts to share using phonetic decision trees
- Using phonetic questions to build a phonetic decision tree
- State splitting - using the likelihood of parent and children states to choose the optimal splitting question
- Mixing up - transforming an HMM-based system based on Gaussian distributions to one based on mixtures of Gaussians
This page maintained by Steve Renals.
Last updated: 2019/01/31 21:43:10UTC
