Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!doc.ic.ac.uk!agate!howland.reston.ans.net!ux1.cso.uiuc.edu!vixen.cso.uiuc.edu!uwm.edu!caen!batcomputer!ghost.sm.dsi.unimi.it!univ-lyon1.fr!scsing.switch.ch!bernina!eeh.ee.ethz.ch!bieler
From: bieler@eeh.ee.ethz.ch (Kaspar Bieler)
Subject: HMM for Alarminterpretation
Message-ID: <CAI7D6.I4@bernina.ethz.ch>
Sender: news@bernina.ethz.ch (USENET News System)
Organization: ETH Zurich, Switzerland
Date: Wed, 21 Jul 1993 07:27:05 GMT
Lines: 36

Hi netters,

I want to use HMM for alarminterpretation. There are 18 different alarm messages.
There are 3 typical classes of faults. For each fault class I have a set of
messages belong to them.

A typical set of such faults belong to class A is for example:

1 4 5 6 8 11 0
2 4 5 6 8 11 0
1 4 5 6 7 8 11 0
1 1 4 4 5 5 6 8 11 11 12 13 0
3 3 4 4 5 5 6 7 8 11 12 13 13 0 
...

I train my HMM with this vectors.

Class B und C are trained similarly.

To my problems:

The length of the vectors are very different. In the above example from the length
7 to 14. What a type of HMM (Bakis, Sphinx-Modell) do you suggest for? Should I
bring in tides and where?

Another problem is the number of vectors. There are e.g. 30 vectors in class A,
25 in class B but only 10 in class C. I try to use the vectors in class C twice
but the result wasn't better. Any suggestions are welcomed. 

Are there any papers dealing these problems?


Thanks in advance, Kaspar



