Speech Recognition Engines

Nested
HTK: Using external coefficients vectors problem.
User: iloes
Date: 10/12/2010 1:40 pm
Views: 12780
Rating: 7

Hi!

I'm trying to develop a recognition system pitch-based. So, I need to use external coefficients vectors proceeding from a pitch detector that I've programmed. I use a MatLab function to make a .htk file format where I append pitch coefficients to the head. If I use "HList -h -S train.scp", for example, I obtain for one .htk file:

------------------------------------- Source: ns5 ------------------------------

  Sample Bytes:  200      Sample Kind:   USER

  Num Comps:     50       Sample Period: 10000.0 us

  Num Samples:   1        File Format:   HTK

------------------------------------ Samples: 0->-1 ----------------------------

0:      -1.067   2.653   2.653   2.653   2.653   2.653   6.580   6.580   6.580

15.101

        15.101  15.101  15.101  15.101  15.101   6.580   6.580   6.580   2.653

 2.653

        -1.067  -2.828  -4.588  -4.588 -11.079 -11.079 -11.079 -11.079 -11.079

-7.923

        -6.256  -4.588   2.653   2.653   2.653   2.653   2.653   2.653   2.653

 2.653

        -4.588  -1.067  -4.588  -4.588  -4.588  -4.588  -7.923 -16.207 -16.207 -

24.492

----------------------------------------- END ----------------------------------


I established 10ms to sample period because my pitch detector evaluate pitch from 10ms length samples (I don't know if this is correct). Vector is composed by 50 coefficients and I defined sample kind as USER.

When I use HERest I obtain the next error for every .htk file corresponding to one utterance:

ERROR [-7324] Bad data or over pruning.

It isn't a matching problem. I think that it isn't a threshold problem either, because I relaxed pruning thresholds in the interval 0 to 10000... Probably the problem is that I have only one sample with 50 components but, if I do it backwards I would have 50 samples with just 1 component per sample so, how could I augment samples and components if I only give a value (corresponding to the pitch in Mel scale) per 10ms segment size?

How must I config the config file if I have defined files with coefficients vectors? For example, I don't need SOURCEKIND but TARGETKIND exists because I have defined .htk files yet (I don't need to use HCopy, I have equivalent .mfc HTKBook example files).

Thank you very much.

--- (Edited on 10/12/2010 1:40 pm [GMT-0500] by iloes) ---

Re: HTK: Using external coefficients vectors problem.
User: iloes
Date: 10/15/2010 8:30 am
Views: 97
Rating: 6

I achieved to give a solution to the problem. It was related, as I supposed, to the number of coefficients. Finally, I inverted the coefficient vector (using only one coefficient per sample) and after I augmented the number of features including energy, delta and acceleration coefficients. It works this way until the end without any problem.


Thanks.

--- (Edited on 10/15/2010 8:30 am [GMT-0500] by iloes) ---

Re: HTK: Using external coefficients vectors problem.
User: aspirant
Date: 10/16/2010 4:43 pm
Views: 121
Rating: 5

hi i also want to use different features for speech recognition like plp, rasta instead of mfcc.can u suggest me hw did u convert into .htk ur own features(matrices in matlab).

 

as i followed the same steps from this site upto step5

after that i used "writhtk function" in matlab through which i computed plp coefficients "in htk format".Instead of mfcc i saved them in

C:\cygwin\home\Administrator\voxforge\train\mfcc\sample1.plp and so on

i have changed the codetrain , train.scp accordingly.

but on running HcompV i m getting error as follows:

Administrator@vishal /home/Administrator/voxforge/auto
$ HCompV -A -D -T 1 -C config_p -f 0.01 -m -S train_p.scp -M hmm0 proto_p
C:\cygwin\HTK\htk-3.3-windows-binary_2\htk\HCompV.exe -A -D -T 1 -C config_p -f
0.01 -m -S train_p.scp -M hmm0 proto_p

HTK Configuration Parameters[10]
  Module/Tool     Parameter                  Value
#                 NUMCEPS                       12
#                 CEPLIFTER                     22
#                 NUMCHANS                      26
#                 PREEMCOEF               0.970000
#                 USEHAMMING                  TRUE
#                 WINDOWSIZE         250000.000000
#                 SAVEWITHCRC                 TRUE
#                 SAVECOMPRESSED              TRUE
#                 TARGETRATE         100000.000000
#                 TARGETKIND           PLP_D_A_Z_0

HMM Def Error: LoadHMMSet: Inconsistent HMM macro name at line 1/col 0/char -1 i
n proto_p
  ERROR [+7050]  HMError:
  ERROR [+2028]  Initialise: LoadHMMSet failed
 FATAL ERROR - Terminating program C:\cygwin\HTK\htk-3.3-windows-binary_2\htk\HC
ompV.exe

--- (Edited on 10/16/2010 4:43 pm [GMT-0500] by aspirant) ---

Re: HTK: Using external coefficients vectors problem.
User: aspirant
Date: 10/16/2010 4:43 pm
Views: 114
Rating: 6

hi i also want to use different features for speech recognition like plp, rasta instead of mfcc.can u suggest me hw did u convert into .htk ur own features(matrices in matlab).

 

as i followed the same steps from this site upto step5

after that i used "writhtk function" in matlab through which i computed plp coefficients "in htk format".Instead of mfcc i saved them in

C:\cygwin\home\Administrator\voxforge\train\mfcc\sample1.plp and so on

i have changed the codetrain , train.scp accordingly.

but on running HcompV i m getting error as follows:

Administrator@vishal /home/Administrator/voxforge/auto
$ HCompV -A -D -T 1 -C config_p -f 0.01 -m -S train_p.scp -M hmm0 proto_p
C:\cygwin\HTK\htk-3.3-windows-binary_2\htk\HCompV.exe -A -D -T 1 -C config_p -f
0.01 -m -S train_p.scp -M hmm0 proto_p

HTK Configuration Parameters[10]
  Module/Tool     Parameter                  Value
#                 NUMCEPS                       12
#                 CEPLIFTER                     22
#                 NUMCHANS                      26
#                 PREEMCOEF               0.970000
#                 USEHAMMING                  TRUE
#                 WINDOWSIZE         250000.000000
#                 SAVEWITHCRC                 TRUE
#                 SAVECOMPRESSED              TRUE
#                 TARGETRATE         100000.000000
#                 TARGETKIND           PLP_D_A_Z_0

HMM Def Error: LoadHMMSet: Inconsistent HMM macro name at line 1/col 0/char -1 i
n proto_p
  ERROR [+7050]  HMError:
  ERROR [+2028]  Initialise: LoadHMMSet failed
 FATAL ERROR - Terminating program C:\cygwin\HTK\htk-3.3-windows-binary_2\htk\HC
ompV.exe

--- (Edited on 10/16/2010 4:43 pm [GMT-0500] by aspirant) ---

Re: HTK: Using external coefficients vectors problem.
User: kmaclean
Date: 10/16/2010 6:43 pm
Views: 225
Rating: 4

Cross posted here: unable to run HcompV

 

--- (Edited on 10/16/2010 7:43 pm [GMT-0400] by kmaclean) ---

Re: HTK: Using external coefficients vectors problem.
User: Chua
Date: 3/1/2011 6:07 am
Views: 163
Rating: 6

Hi,

 

I am also thinking of using other feature vectors other than mfcc. but i am not sure on how to implement plp.


i am very new to this and my project supervisors have basically left me to get a feel of this. correct me if i am wrong, i can change it under the targetkind to plp right.

 

what kind of changes do i have to make to my wave files and codetrain?

--- (Edited on 3/1/2011 6:07 am [GMT-0600] by Visitor) ---

Re: HTK: Using external coefficients vectors problem.
User: Chua
Date: 3/1/2011 6:10 am
Views: 689
Rating: 4

I meant other than obviously renaming all mfcc related folders and extensions to plp. i read that some users have reconfigured that deltas etc, are these required?

 

thank

--- (Edited on 3/1/2011 6:10 am [GMT-0600] by Visitor) ---

Re: HTK: Using external coefficients vectors problem.
User: eslam
Date: 7/20/2012 7:57 am
Views: 365
Rating: 6

Al Salam Alikoum,

I also work with new feature extraction method for phone recognition but the correct percentage is very small as I get 25%  and the big problem is the accuracy is negative .

any one can help me to improve the result and give accuracy with suitable value

--- (Edited on 7/20/2012 7:57 am [GMT-0500] by Visitor) ---

Re: HTK: Using external coefficients vectors problem.
User: Nitish
Date: 9/5/2012 6:06 am
Views: 948
Rating: 6

Hi all,

Im basically working on the same idea. Here's what I am doing. I use HTK to convert my .wav files into .mfc files. Then I use HList to convert .mfc files to text files. I get the pitch values from Praat. Im trying to append the pitch values to the correspinding mfcc vectors. Here's my problem. Say, a wav file is of 79 seconds duration. I have set TARGETRATE=100000 and WINDOWSIZE=200000, so (please correct me if im wrong)  I should be getting like 7900 vectors or so in my .mfc files. Instead, I get 3950 vectors ( half of what I expected). I am sure I am missing something. Please help!

--- (Edited on 9/5/2012 6:06 am [GMT-0500] by Visitor) ---

Re: HTK: Using external coefficients vectors problem.
User: Visitor
Date: 7/31/2014 2:58 am
Views: 2219
Rating: 3

Hi..

Can u kindly tell me how to append the pitch etc to mfcc features using HTK..

--- (Edited on 7/31/2014 2:58 am [GMT-0500] by Visitor) ---

PreviousNext