Acoustic Model Discussions

Flat
Need help MFCC_0_D_A vs MFCC_0_D_N_Z
User: xxx
Date: 8/27/2008 1:38 am
Views: 10288
Rating: 14
Hi

I want to know what is the differece between MFCC_0_D_A and MFCC_0_D_N_Z and which one is better?

Also if i use MFCC_0_D_N_Z i can't perform HCopy directly usin this option and therefore i will have to use
something like MFCC_0_D and then later change to MFCC_0_D_N_Z. But when it comes to testing i get a error saying Incompatible sample kind MFCC_D_0 vs MFCC_D_N_Z_0
How do i overcome this?

Also once i have completed making my HMM model with MFCC_0_D_N_Z, I am NOT able to run HVite live aswell. I receive the following error
READY[1]>
  ERROR [+6320]  OpenAsChannel: cannot zero mean within buffer
  ERROR [+6316]  OpenBuffer: OpenAsChannel failed
  ERROR [+3250]  ProcessFile: Config parameters invalid
 FATAL ERROR - Terminating program HVite


my config file for recognition looks as follows
=========================================================================================
                    #config2
# Waveform Capture
SOURCERATE = 625.0    # = 16kHz
SOURCEKIND = HAUDIO
SOURCEFORMAT = HTK

#Conversion parameters of the input signal
AUDIOSIG = -1        # Negative value => key press control


#USESILDET = T
#MEASURESIL = F
#OUTSILWARN = T
#SILENERGY = 55.0

ENORMALISE = F
TARGETKIND = MFCC_0_D_N_Z        #Identifier of the coefficients to use
TARGETRATE = 100000.0                # = 10 ms = frame periodicity
WINDOWSIZE = 250000.0                # = 25 ms = length of time frame
USEHAMMING = T                            #Use of Hamming windowing function
PREEMCOEF = 0.97                        #Pre-emphasis coefficient
NUMCHANS = 26                                #Number of filterbank channels
CEPLIFTER = 22                            #Length of cepstral liftering
NUMCEPS = 12                                #Number of MFCC coeffs
============================================================================================

config for wav is when creating the MFCC_0_D_N_Z model (for HCOpy)
=====================================
# Coding parameters
TARGETKIND = MFCC_0_D
TARGETRATE = 100000.0
SAVECOMPRESSED = T
SAVEWITHCRC = T
WINDOWSIZE = 250000.0
USEHAMMING = T
PREEMCOEF = 0.97
NUMCHANS = 26
CEPLIFTER = 22
NUMCEPS = 12
ENORMALISE = F
SOURCEFORMAT = WAV
SOURCEKIND = WAVEFORM
TARGETFORMAT = HTK
============================================

later this is change to (to be used in HERest)
==============================================================================
# Coding parameters
TARGETKIND = MFCC_0_D_N_Z
TARGETRATE = 100000.0
SAVECOMPRESSED = T
SAVEWITHCRC = T
WINDOWSIZE = 250000.0
USEHAMMING = T
PREEMCOEF = 0.97
NUMCHANS = 26
CEPLIFTER = 22
NUMCEPS = 12
ENORMALISE = F ======================================================


thanks heaps !!!

--- (Edited on 8/27/2008 1:38 am [GMT-0500] by Visitor) ---

Re: Need help MFCC_0_D_A vs MFCC_0_D_N_Z
User: kmaclean
Date: 8/27/2008 9:08 am
Views: 1850
Rating: 10

Hi triple x,

> Also if i use MFCC_0_D_N_Z i can't perform HCopy directly usin this

> option and therefore i will have to use something like MFCC_0_D and then

> later change to MFCC_0_D_N_Z. But when it comes to testing i get a error

> saying Incompatible sample kind MFCC_D_0 vs MFCC_D_N_Z_0

>How do i overcome this?

You need to do the converstion from MFCC_0_D to MFCC_0_D_N_Z in the proto file and with the HCompV command, not by using HERest...  See this post MFCC_D_N_Z_0 format, and see Step 6 - Creating Flat Start Monophones of the VoxForge acoustic model creation tutorial.

Ken

--- (Edited on 8/27/2008 10:08 am [GMT-0400] by kmaclean) ---

Re: Need help MFCC_0_D_A vs MFCC_0_D_N_Z
User: NancyZxll
Date: 5/29/2012 10:42 am
Views: 1361
Rating: 10

Hello Ken,

    I met the same mistack with Triple x.

    I firstly used config file as:

----------------------------------------------

SOURCEFORMAT = WAV

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

-------------------------------------------

to run HCopy and got the MFCC test file.

Then I use HVite to recognise the test data.

It errors like:  

ERROR [+3231]  ProcessFile: Incompatible sample kind MFCC_D_A_0 vs MFCC_D_N_Z_0

 

=======================================

Then I tried to use "TARGETKIND = MFCC_0_D_N_Z" in the config and ran the HCopy command again.

The error is:  

ERROR [+1019]  SetConfParms: incompatible TARGETKIND=MFCC_0_D_N_Z for coding

 

Neither method could work..

Could you please help me?

--- (Edited on 5/29/2012 10:42 am [GMT-0500] by NancyZxll) ---

Re: Need help MFCC_0_D_A vs MFCC_0_D_N_Z
User: Meysam
Date: 8/16/2014 3:10 am
Views: 4052
Rating: 5

Hi
you must use config file in HVite command. I have this problem with this form use of HVite:

HVite -T 1 -S Test\test.scp -m -H hmm4\hmmdefs -i recout.mlf -w wdnet voxlexi tiedlist

but when I use HVite this form it's ok:

HVite -T 1 -C config -S Test\test.scp -m -H hmm4\hmmdefs -i recout.mlf -w wdnet voxlexi tiedlist

 

my config file was:

----------------------------------------

TARGETKIND = MFCC_0_D_N_Z

TARGETRATE = 100000.0

SAVECOMPRESSED = T

SAVEWITHCRC = T

WINDOWSIZE = 250000.0

USEHAMMING = T

PREEMCOEF = 0.97

NUMCHANS = 26

CEPLIFTER = 22

NUMCEPS = 12 

--------------------------

--- (Edited on 8/16/2014 3:10 am [GMT-0500] by Meysam) ---

PreviousNext