-- Main.matthewa - 04 Aug 2006

I think we are going to have fun with silence in this framework. If you have a look at the example alignment (after only 4 iterations over only 100 files) you can see a nice pattern in the silence s86->s83->s86 etc. However in the speech (Computers) the s83 state is appearing again for completely different bits of waveform.

I was thinking maybe we will need to bootstrap a silence model in some way.

Another problem is that during training I get these sort of messages:

WARNING [-2326] UpdateTrans : Model 1[b]: no transitions out of state 16 in HERest

WARNING [-2330] UpdateVars : Model 1[b]: no use of variance 16.1.1 in HERest

WARNING [-2330] UpdateMeans : Model 1[b]: no use of mean 16.1.1 in HERest

etc. I assume we are losing states. If you keep running HERest it after aboiut 9 iterations the whole thing pretty much implodes.

  • example_state_alignment.gif:
Topic attachments
I Attachment Action Size Date Who Comment
gifgif example_state_alignment.gif manage 24.7 K 04 Aug 2006 - 16:13 Main.matthewa  
Topic revision: r1 - 04 Aug 2006 - 16:13:32 - Main.matthewa
This site is powered by the TWiki collaboration platformCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback
This Wiki uses Cookies