SP Module 10 Connected Speech & HMM Training

From subword units to n-grams: hierarchy of models

Defining a hierarchy of models: we can compile different HMMs to create models of utterances

s11243611272022

s11201111272022

s11241211272022

s11250411272022

s11270611272022

s11275311272022

s11285111272022

s11295011272022

We can do some pruning, remove some tokens while proceeding, reduce computation cost (Maybe Heuristic is also can be helpful in such case.)

Conditional independence and the forward algorithm

We use the Markov property of HMMs (i.e. conditional independence assumptions) to make computing probabilities of observation sequences easier

s14114011282022

s14145911282022

HMM training with the Baum-Welch algorithm

HMM training using the Baum-Welch algorithm. This gives a very high level overview of forward and backward probability calculation on HMMs and Expectation-Maximization as a way to optimise model parameters. The maths is in the readings (but not examinable).

s14164611282022

s14293011282022

Origin: Module 10 – Speech Recognition – Connected speech & HMM training Translate + Edit: YangSier (Homepage)