On 8 Apr 2014, at 20:10, Majid Namnabat <maj.nam@xxxxxxxxx> wrote: In HTS training when copying mono-phone HMMs to full context HMMs, and then HERest (embedded re-estimation) of full context HMMs are done, all question features are considered in full context phoneme classification. For example forward and backward distance in word, phrase , ... are considered ! My belief is these features are redundant that cause sparsity and so full context HMM re-estimated with only rare instances and so these low-training have straight affection in tying and clustering after that. Namnabat, yes, when training *untied* context-dependent models, they will be very poorly estimated due to the sparsity in the data. However, the reason we want these models is only so that a decision tree can be grown, which will cluster them. After that, *clustered* models will be trained, and so there will be sufficient data per model. You are certainly right in saying that the decision tree is grown on the basis of very poorly estimated untied models. But, this is not as bad as it sounds: the tree only needs to use these poor models in order to discover which clusters of them have similar parameters, so it can tie them together. Discovering this similarity is the most important thing; it is less important whether the models are actually any good for speech synthesis (or recognition). In some "recipes", the clustered models are trained for several iterations, then untied, trained whilst untied for one iteration, and re-clustered again. This might be repeated several times. This process may mitigate the fact that the very first decision tree was grown on the basis of very poor models. Simon
--
Prof. Simon King Professor of Speech Processing & Director of the Centre for Speech Technology Research University of Edinburgh,UK www.cstr.ed.ac.uk |
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.