In running the example TIMIT phoneme recognition demo I found two errors in the file TimitGetFiles.py which invokes HTK toolkit HCopy function to transform TIMIT audio files into feature files for CNTK:
Line 44 has a syntax error. It looks like Matlab syntax somehow crept into this python file:
I don't know the correct python syntax; I just commented out lines 44-46.
- if !hcopyScript or !cnScriptFp or !cnOutputScriptFp or !fileListFp:
hcopyScriptFp.write(fullFile+'.nst ' + featFile+'.fbank_zda\n')
My TIMIT database has its NIST-format files named with extension .wav. I had to change the .nst in line 55 to .wav. Somebody who knows python could write code to automatically detect the extension.
PS: after these changes I was able to run the 25 epoch TIMIT phoneme training in 50 seconds using a GTX 660 Ti card, which appeared to operate at 50 to 60 % load on the GPU. CPU was i7 3770K.