Hi,
In forward-backward training, no matter embedded or non-embedded training, the state alignement would be changed.
So maybe, in the previous steps, each state have enough data, and when it comes to embedded training of the clustered model set, state boudary change results in a bad data error.
12 might be too much for some phones, considering 5ms frame length.
I don't know if my understanding is right. :-)
--
Na Xingyu (那兴宇)
Beijing Institute of Technology
naxy(at)bit.edu.cn
asr.naxingyu(at)gmail.com
naxingyu at {facebook, twitter, linkedin}
At 2011-10-06 17:57:50,"Lee Siu Wa Yvonne" <swylee@xxxxxxxxxxxxxxxxx> wrote:
>Thanks for your reply.
>
>My problem is that the training process stopped during embedded re-estimation (clustered) after tree clustering, but not in previous steps (e.g. training monophone models or full-context models).
>
>If this is the case for models having more states than the available frames in training files, should the script stops earlier, e.g. HERest for monophone or full-context phone models?
>
>Regards,
>LEE Siu Wa
>
>________________________________
>
>From: Simon King [mailto:Simon.King@xxxxxxxx]
>Sent: Thu 10/6/2011 4:17 PM
>To: hts-users@xxxxxxxxxxxxxxx
>Subject: [hts-users:03048] Re: Increasing the no. of statea
>
>
>
>On 6 Oct 2011, at 09:01, Lee Siu Wa Yvonne wrote:
>
>> I have changed $nState previously from 5 to 12, but the training process stopped during embedded re-estimation (clustered) with the error message below:
>
>> No path found in beta pass
>
>You may have created models that have more states than there are frames in some of your shorter training files, which means there is no possible state alignment for those particular files. That's what the 'no path' message could be telling you.
>
>Simon
>
>
>--
>The University of Edinburgh is a charitable body, registered in
>Scotland, with registration number SC005336.
>
>
>
>
>“1991-2011 - Creating Growth, Enhancing Lives. Commemorating A*STAR’s 20 Years of Science, Technology and Research in Singapore.â€
>
>Institute for Infocomm Research disclaimer: "This email is confidential and may be privileged. If you are not the intended recipient, please delete it and notify us immediately. Please do not copy or use it for any purpose, or disclose its contents to any other person. Thank you."
>