VoxForge
VoxForge
from this post by kalmanbucy:
Re: Error [+6510] LOpen: Unable to open label file downsampled.lab
In the "A Tutorial Example" of HTKbook, we need to put .lab files with those .mfc in the feature folder as we create flat start monophones. Then we execute "HERest" three times to train the model. After that we add a model for "sp" in the step 7 to fix silence model, and use "HLEd" to generate a new trascription file for all phones plus "sp", called phones1, right. Then we need to train the models twice more. However, an issue comes up. The .lab files under .data/train/feature remains old, i.e. no "sp" was added to these .lab files while silence model was fixed.If we follow the htk tutorial to execute this below:
HERest -C config/config1 -I labels/phones1.mlf -t 250.0 150.0 1000.0 -S train.scp -H hmms/hmm5/macros -H hmms/hmm5/hmmdefs -M hmms/hmm6 lists/monophones1
A warning will pop up:
Pruning-On[250.0 150.0 1000.0]
WARNING [-2331] UpdateModels: sp[25] copied: only 0 egs
in HERestp:
It means model can't be updated due to little training data, because .lab will be searched in the process of training. Have you ever meet this trouble before?
Hi kalmanbucy:
You should have created two mlf files (Multi-Label Files containing many "label" files) in Step 4 of the HTK Tutorial . One with sp pauses inserted and one without. The following is taken from Step 4 - Creating the Transcription Files of the VoxForge tutorial:
Next, we need to create a second phones1.mlf file (which will include short pauses
(“sp”) after each word phone group). First create the
mkphones1.led in your
'voxforge/manual' folder as follows:
|
EX |
(note: remember to include a blank line at the end of this script)
Then run the HLEd command again from your 'voxforge/manual' folder as follows:
| $HLEd -A -D -T 1 -l '*' -d dict -i phones1.mlf mkphones1.led words.mlf |
Which creates the
phones1.mlf
file.
Ken
Hi Ken,
Thank you for your help.
I think this issue was solved by strictly changing the "*/2..lab" to "*/S0002.lab" which shares the same prefix with those speech files, then it could search the .mlf successfully. Looks like no .lab files are needed additionally.
But right now, I'm confused with the anthoer bad performance in the final recognition after step 7.
I record 87 utterances with HSLab, then convert them to .mfc files. Then follow the instructions in HTKtuotorial. Finishing "Step 7 - Fixing the Silence Models", I checked the result right now by
"HVite -H hmms\hmm7\macros -H hmms\hmm7\hmmdefs -S test.scp -l * -i results\recout_step7.mlf -w wdnet -p 0.0 -s 5.0 dicts\dict1 lists\monophones1
HResults -I labels\testwords.mlf lists\monophones1 results\recout_step7.mlf"
The result is
====================== HTK Results Analysis=====
Date: Tue Mar 17 22:50:40 2009
Ref : labels\testwords.mlf
Rec : results\recout_step7.mlf
------------------------ Overall Results --------------------------
SENT: %Correct=0.00 [H=0, S=2, N=2]
WORD: %Corr=0.00, Acc=-8.33 [H=0, D=4, S=8, I=1, N=12]
===========================================
But I had a very successful recognition result as I start over the whole process with some other .mfc files from a friend. Unfortunately, speech files are lost. According to the comparision, I personally doubt that the reason is associated with the coding process converting voices to .mfc parameter files. Both HSLab and training process are running on Windows XP. Here is my config1 content:
TARGETKIND = MFCC_0_D_A
TARGETRATE = 100000.0
SAVECOMPRESSED = T
SAVEWITHCRC = T
WINDOWSIZE = 250000.0
USEHAMMING = T
PREEMCOEF = 0.97
NUMCHANS = 26
CEPLIFTER = 22
NUMCEPS = 12
ENORMALISE = F
SOURCERATE = 625
Could analyze why I get the bad result with my own utterance files? Thanks.
Hi kalmanbucy,
>But I had a very successful recognition result as I start over the whole
>process with some other .mfc files from a friend.
Because of this, it sounds like you did something wrong in the coding process.
>I personally doubt that the reason is associated with the coding process
>converting voices to .mfc parameter files. [...]
Did you try looking at the headers of your files using HList (are they corrupted)? Did you record your data at a different SOURCERATE than what is in you config file?
Ken