Hi nsh,

Thanks for the link.

This is a great resource! And it's really nice that they supply uncompressed audio (wav format) and are using the GPL license.

The one downside with them is that the speech files are unsegmented (at least the one I looked at ...).  We would still need to segment the speech into 10-15 word sentences so that Sphinx or HTK could create acoustic models from them. 

thanks!

Ken 

--- (Edited on 10/21/2007 8:14 pm [GMT-0400] by kmaclean) ---

There is transcription at least for some of dialogs. it's inside their format, but I suppose it can be extracted easily:

 SUZ:   +, in order for this one unit of weight ^%mov:"GRAVITY"_66919_68750^    
        to be pulled it takes one Newton . ^%mov:"GRAVITY"_68750_71582^

 numbers here is a time I suppose

--- (Edited on 10/22/2007 2:42 am [GMT-0500] by nsh) ---

Hi nsh, 

My mistake ... you're right!

They have a portion of the Switchboard corpus that has time alignments, which looks like this:

sw2019A-ms98-a-0001 0.000000 0.550000 [silence]
sw2019A-ms98-a-0001 0.550000 0.630000 [noise]
sw2019A-ms98-a-0001 0.630000 1.131000 [silence]
sw2019A-ms98-a-0002 1.131000 2.421000 [silence]
sw2019A-ms98-a-0002 2.421000 2.601000 [silence]
sw2019A-ms98-a-0002 2.601000 2.699625 [silence]
sw2019A-ms98-a-0003 2.699625 3.049625 [silence]
sw2019A-ms98-a-0003 3.049625 3.219625 uh
sw2019A-ms98-a-0003 3.219625 3.349625 do
sw2019A-ms98-a-0003 3.349625 3.489625 you
sw2019A-ms98-a-0003 3.489625 3.609625 have
sw2019A-ms98-a-0003 3.609625 3.679625 a
sw2019A-ms98-a-0003 3.679625 3.929625 pet

It should be easy to create a script to segment the audio using this.

I need to take some time to go over this in detail. 

thanks,

Ken 

--- (Edited on 10/22/2007 10:56 am [GMT-0400] by kmaclean) ---