[hts-users:00375] Re:

[Keiichi Tokuda <tokuda@xxxxxxxxxxxxxxxx>]

From: "Ossama AbdelHamid" <ossama_a@xxxxxxxxx>
Subject: [hts-users:00374] 
Date: Mon, 17 Jul 2006 19:53:02 GMT
Message-ID: <20060717195302.9624.qmail@xxxxxxxxxxxxxxx>

> I have noticed that in the HTS version hts-2.0RC2, in the
> training first you cluster the models then untie them then
> recluster again. What's the gain of this untieing-reclustering
> process?

It is to solve a chicken and egg problem.  Clustering is done
based on the sufficient statistics collected in the embedded
training stage just before the clustering.  However, it is based
on monophone models (or non-clustered context-dependent models).
To collect better statistics, we can use the clustered models,
and then re-cluster them.  The process (clustering, embedded
training, and untying) can be repeated to improve the model
accuracy.

Keiichi Tokuda
tokuda@xxxxxxxxxxxxxxxx
http://kt-lab.ics.nitech.ac.jp/~tokuda/