Hi Ken,
On 32bit version of Ubuntu HTK compiled without a problem. Then I followed the tutorial "Automated Audio Segmentation Using Forced Alignment" without a hitch up to step 6. In Audacity I see a considerable misalignement between audio and labels. For example in the phrase "Chapter Four, Part One" the audio corresponding to label "part" contains sounds "part one", and the audio corresponding to label "one" contains only silence. Is it a normal misalignement or am I doing something wrong?
Thanks a lot,
--Sergey

Oops. Looks like I have posted the message in a wrong thread and I don't see any controls how to move it. Sorry for the mess.
--Sergey

Hi Sergey, 

>Looks like I have posted the message in a wrong thread

It's a bit confusing, but this forum is OK - since your question relates to stuff on the dev page.  I really should have a comments page on the Automated Audio Segmentation page... but never got around to it :)

Ken 

 

Hi Sergey,

>Is it a normal misalignement or am I doing something wrong?

It all depends on the acoustic model - the better the AM, the better the alignments.  Are you using the current stable VoxForge acoustic model or one of the nightly builds?

Remember for audio segmentation, we don't need perfect alignment, just one good enough to pick out the silences.

Ken 

Hi Sergey,

I fixed the link to the audiobook used in the automated segmentation howto:

historyofengland01ch04_01_macaulay.wav 04-Mar-2007 14:32 165M

Look at this audio with the labels I generated (audacityLabelTrack.txt) in Audacity ... are yours any worse/better than these?

Ken 

 

> Are you using the current stable VoxForge acoustic model or one of the nightly builds?
I used the nightly build, because when I click the click to the "current stable release" I get "Error 404: NOT FOUND!"

> ...are yours any worse/better than these?
Oh, mine are much worse. Yours are perfect compared to mine.

Hi Sergey,

>Now the alignment look good!

Now that's interesting ... for the longest time, the nightly build AM performed better than the current stable release ... I'll have to look in to this.

thanks for your feedback! 

Ken 

I have found a newer version of acoustic model which also works very good.
http://www.repository.voxforge1.org/downloads/Main/Tags/Releases/0_1_1-build726/

Just a note to let you know that I removed some noisy audio from the corpus last week (actually, I the audio is still there, but Iremoved their entry from the master_prompts files) - this seems to have cleared up this problem.  I tried the segmentation process with the May 27, 2008 Nightly acoustic model build, and it now segments properly.

Ken