Speech Recognition Engines

Re: Number recognition
User: Visitor
Date: 10/19/2010 6:22 pm
Views: 192
Rating: 1

I can confirm that using the -argfile argument works perfectly.

In the chapter "Converting the sendump and mdef files", multiple ways of obtaining the mixture_weights file are described. The mixture_weights file is not available for TIDIGTS in cmusphinx-pocketsphinx-extra.tar.gz, so the SphinxTrain Python modules have been build from SpinxTrain-1.0 and sendump.py is run like this:

./sendump.py /home/smith388/project/tidigits/sendump /home/smith388/project/tidigits/mixture_weights

The following error message is printed:


Traceback (most recent call last):
  File "./sendump.py", line 26, in <module>
    header = readstr(sendump)
  File "./sendump.py", line 13, in readstr
    nbytes = struct.unpack('I', fh.read(4))[0]
struct.error: unpack requires a string argument of length 4


But at the following location:


a file named mixture_weights is stored. Is it possible to just copy that file to my custom tidigits folder without problems?

--- (Edited on 10/19/2010 6:22 pm [GMT-0500] by Visitor) ---

Re: Number recognition
User: nsh
Date: 10/21/2010 10:09 am
Views: 2945
Rating: 1


sendump in pocketsphinx/tidigits uses new format which is not supported by sendump.py. Unfortunately it's not easy to convert it. It will take some time to write a script.

I think the easiest way to obtain mixture_weights is to ask for this file on cmusphinx-devel mailing list.


--- (Edited on 10/21/2010 19:09 [GMT+0400] by nsh) ---