VoxForge
VoxForge
Hi Ken and All,
Thanks for all the helpful hints in the forums. I could solve quite some errors but it seems I don't know what to do about this one.
[By now I've been trice through the process of building a single-word (10 only) speech recogniser with HTK using the HTK Book and the Voxforge Tutorial. The first time my recognition results were 100% but even when running the recogniser offline it only recognised two of these ten words even when different waves were loaded that did not contain these two words. Anyhow, I felt there was something wrong so I started over again and again.]
This time, in step 7 of the Voxforge Tutorial, when executing the following:
laptop:~$ HERest -C config -I phones0.mlf -t 250.0 150.0 1000.0 -S train.scp -H hmm5/macros -H hmm5/hmmdefs -M hmm6 monophones1
I get this warning:
Pruning-On[250.0 150.0 1000.0]
WARNING [-2331] UpdateModels: sp[20] copied: only 0 egs in HERest
I'm not sure how serious it is and whether I should ignore or solve it. Does anyone know what the solution might be?
I have a single-word grammar with ten words. These ten words are trained with 345 wav-files containing one word each. It seems quite problematic if "0 egs" out of 345 wav-files can be processed.
I converted my files from 44100Hz to 16000Hz, because I read that for the SOURCERATE to be 625.0 16KHz is the right sampling rate. Now earlier, when executing HCopy -T 1 -C config -S codetr.scp to create the *.mfc's I used the configuration parameter TARGETKIND = MFCC_0_D_A, although the HTK Tutorial suggests to use TARGETKIND = MFCC_0 in step 5 of the HTK Tutorial. However, when using TARGETKIND = MFCC_0 and one step further executing:
HERest -C config -I phones0.mlf -t 250.0 150.0 1000.0 -S train.scp -H hmm0/macros -H hmm0/hmmdefs -M hmm1 monophones0
almost all of my wav-files got the following error:
WARNING [-7324] StepBack: File /*.mfc - bad data or over pruning in HERest
So in general, there is something wrong with my wav-files. Are they too short (min = 0.43sec, max = 1.66sec, usually 0.8sec)?
I'd appreciate any help I can get! Cheers.
Katie
--- (Edited on 9/6/2009 8:20 pm [GMT-0500] by Visitor) ---
>Pruning-On[250.0 150.0 1000.0]
> WARNING [-2331] UpdateModels: sp[20] copied: only 0 egs in
>HERest
>I'm not sure how serious it is and whether I should ignore or solve it.
See my post in this thread: Create hmm11 - usually this means that you don't have enough speech for a particular phone, and can be ignored... However, this is the 'sp' phone. I think your problem is related to this statement: "These ten words are trained with 345 wav-files containing one word each."
The "sp" phone is the short-pause interword phone... if all you have are single word recordings, then you will not have any short-pauses. See Step 7 - Fixing the Silence Models for a short description of how the short-pause model is created. You might try recording a few sentences (multiple words) and see if this helps.
>Are they too short (min = 0.43sec, max = 1.66sec, usually 0.8sec)?
I really don't know - have you tried with longer wav files? Do you get the same error? You say "almost all of my wav-files got the following error:" - are the ones that do not get an error longer wav files?
>WARNING [-7324] StepBack: File /*.mfc - bad data or over
>pruning in HERest
look up the "-t" parameter (from -t 250.0 150.0 1000.0) in HTK book, I think these can be adjusted so that you get higher training thresholds...
Ken
--- (Edited on 9/9/2009 11:32 pm [GMT-0400] by kmaclean) ---
I got error like this:
HERest -A -t 600.0 -I batiktypephone.mlf -S AllTrainData1.scp -H hmm1/hmmdefs -M hmm1 model.list
Pruning-On[600.0]
WARNING [-7324] StepBack: File C:/Program Files/MATLAB71/work/Skripsi_065314012/data_1/1_1.lt - bad data or over pruning
in HERest
...
WARNING [-2331] UpdateModels: tambal[1] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: truntum[2] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: ceplok_dempel[3] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: nitik_cengkeh[4] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: kawung_picis[5] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: sil[6] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_barong[7] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: pari_sewuli_seling_pamor[8] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_tuding[9] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: ceplok_kece[10] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: kawung_galar[11] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: kawung_kembang_cempoko_hitam[12] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_pancing[13] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_curigo[14] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_kusumo_ukel[15] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_kawung_picis[16] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_kembang_sawut_cecek[17] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_klitik[18] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: lereng_kusuma[19] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_keris[20] copied: only 0 egs
in HERest
WARNING [-2331] UpdateModels: parang_kunci[21] copied: only 0 egs
in HERest
HElp for this error...thank
--- (Edited on 5/18/2010 11:45 pm [GMT-0500] by prina) ---
> WARNING [-7324] StepBack: File C:/Program
>Files/MATLAB71/work/Skripsi_065314012/data_1/1_1.lt - bad
>data or over pruning
Did you check the '1_1.lt' file for bad data?
>WARNING [-2331] UpdateModels: tambal[1] copied: only 0 egs
As I stated in the post above:
See my post in this thread: Create hmm11 - usually this means that you don't have enough speech for a particular phone, and can be ignored...
Why are you using more than two letters for your phones (e.g. tambal, ...)?
--- (Edited on 6/9/2010 9:06 pm [GMT-0400] by kmaclean) ---