[Chiara]

Thanks, that actually solved the problem. Nevertheless, endpointing did not help with the alignment: the script runs fine now, but still throws the beam search threshold error. I tried increasing the values again, and adding another 40 training sentences, but no success yet. Any suggestions about what could make the models good enough with only this little training, or what may be causing the problem?

[Chiara]

Following up on your suggestion: I tried this and I am now getting a different error at the very beginning of the execution,

CALCULATING VARIANCE FLOOR using HCompV
ERROR [+6310] OpenParmChannel: cannot open Parm File /mnt/courses.homes/s1345717/Documents/ss/mfcc/arctic_a0001.mfcc

and a couple of other warnings:

ERROR [+6313] OpenAsChannel: OpenParmChannel failed
ERROR [+6316] OpenBuffer: OpenAsChannel failed
ERROR [+2050] LoadFile: Config parameters invalid

Not sure of why it cannot find the files (also the path there is wrong, the mfccs are not located in the “ss” folder!). Where in the script is the program redirected there? Why does it only happen with the modified version of the script, while program runs fine with the normal do_alignment even after having endpointed the wavs?
Thank you.

[Chiara]

When running the modified do_alignment script, training the system on half of the data (only the “a” section of Arctic) and testing it on all data, I keep getting the following error:

“try a larger beam search thresh for HVite.”

I have increased the beam threshold(s) (P1,P2,P3) by a factor of 10 and 100, but I keep getting the same. I am not sure of how to change the script so to make the alignment possible?

[Author]

Posts