Training Model with another feature vector

[Do Quoc Truong]

Hi,

I completed training model sphinx4 for digits with MFCC's feature, but now I
want to try another feature vector (like PLP...) but I don't know how.

I've searched it but I couldn't find out in anywhere.

Your help is appreciated.

Tks.

[Nickolay V. Shmyrev]

but now I want to try another feature vector

Specifically for PLP you can extract features with sphinx4 feature extraction
tool instead of make_feats.pl:

java -cp lib/sphinx4.jar:lib/batch.jar \
        edu.cmu.sphinx.tools.feature.FeatureFileDumper -name plpFrontEnd \
        -i wav -o feat -ctl etc/db_train.fileids

Only note that sphinx4 PLP extraction by default works with 8khz audio.

As for other feature type, you need to implement a tool to create MFC files
yourself. The format of the MFC file is very simple it's just the number of
floats followed by feature vector values.

[Do Quoc Truong]

Thanks nshmyrev! I can't find it in any document.

I want to compare accuracy by using different feature vector. Do you have any
suggestion?, I appreciated about that.

Thanks for your help!

[Nickolay V. Shmyrev]

Do you have any suggestion?,

Don't bother with that.

[Do Quoc Truong]

I've extracted features with sphinx4 feature extraction tool. but when I run
script/Runall.pl to training I got error below:

WARNING: "init_gau.c", line 264: utt VI-516.story1 too short
Header size field: 0
ERROR: "corpus.c", line 1555: MFCC read of /home/ipman/ivy/NCKH/training/digits/feat/train/VI-20033.nums.mfc failed.  Retrying after sleep...
Header size field: 0
ERROR: "corpus.c", line 1555: MFCC read of /home/ipman/ivy/NCKH/training/digits/feat/train/VI-20033.nums.mfc failed.  Retrying after sleep...
Header size field: 0
ERROR: "corpus.c", line 1555: MFCC read of /home/ipman/ivy/NCKH/training/digits/feat/train/VI-20033.nums.mfc failed.  Retrying after sleep...
Header size field: 0
....

Can you tell me the way to handle it!

Tks.

[Nickolay V. Shmyrev]

Maybe something went wrong. Check that files were actually created, print
their contents with sphinx_cepview tool.

[Do Quoc Truong]

I've check that files, some files just have 4 B, it contained nothing, there
is no problem with audio file.
I ignore that files have error, But when I training I get error below:

ERROR: "backward.c", line 430: Failed to align audio to trancript: final state of the search is not reached
ERROR: "baum_welch.c", line 331: train/VI-277.areacode ignored

I printed VI-277.areacode 's contents:

INFO: main_cepview.c(194): Displaying 10 out of 13 columns per frame
INFO: main_cepview.c(195): Total 32 frames

  6.188  -0.877  -0.241  -0.067  -0.186  -0.076  -0.061  -0.046  -0.025  -0.044 
  6.091  -0.767  -0.252  -0.098  -0.008  -0.094  -0.078  -0.075   0.025  -0.026 
  7.316  -0.784  -0.083  -0.100  -0.205  -0.096  -0.004  -0.014  -0.026  -0.047 
  7.639  -0.796  -0.106  -0.121  -0.176  -0.108  -0.033  -0.038  -0.020  -0.026 
  7.400  -0.944  -0.177  -0.099  -0.049  -0.072  -0.078  -0.077  -0.001  -0.017 
  7.467  -0.830  -0.149  -0.157   0.000  -0.060  -0.037  -0.043  -0.043  -0.038 
  7.518  -0.766  -0.150  -0.161  -0.013  -0.033  -0.054  -0.072  -0.045  -0.031 
  7.491  -0.838  -0.193  -0.170   0.009  -0.095  -0.090  -0.068  -0.024  -0.015 
  7.551  -0.830  -0.187  -0.141  -0.077  -0.119  -0.081  -0.057  -0.034  -0.053 
  7.638  -0.792  -0.092  -0.232  -0.127  -0.063  -0.096  -0.078  -0.029  -0.055 
  7.581  -0.810  -0.069  -0.117  -0.109  -0.082  -0.073  -0.100  -0.021  -0.007 
  7.277  -0.950  -0.145  -0.198  -0.063  -0.044  -0.089  -0.064  -0.025  -0.003 
  7.338  -0.986  -0.131  -0.160  -0.043  -0.066  -0.082  -0.057   0.010  -0.032 
  7.363  -0.918  -0.084  -0.213  -0.030  -0.058  -0.093  -0.033   0.017  -0.058 
  7.472  -0.807  -0.110  -0.204  -0.073  -0.091  -0.077  -0.003   0.002  -0.030 
  7.568  -0.747  -0.172  -0.207  -0.076  -0.119  -0.081  -0.025  -0.019  -0.037 
  7.573  -0.761  -0.186  -0.181  -0.072  -0.067  -0.095  -0.047  -0.008  -0.042 
  7.561  -0.820  -0.173  -0.174  -0.117  -0.085  -0.078  -0.072  -0.001  -0.024 
  7.669  -0.804  -0.263  -0.215  -0.142  -0.074  -0.035  -0.075  -0.040  -0.003 
  7.654  -0.763  -0.287  -0.254  -0.084  -0.010  -0.049  -0.077  -0.062  -0.011 
  7.650  -0.774  -0.253  -0.200  -0.108  -0.058  -0.036  -0.123  -0.068  -0.012 
  7.556  -0.772  -0.164  -0.176  -0.090  -0.083  -0.045  -0.070  -0.038  -0.024 
  7.573  -0.787  -0.158  -0.172  -0.098  -0.082  -0.055  -0.033  -0.068  -0.043 
  7.507  -0.788  -0.130  -0.150  -0.083  -0.094  -0.029   0.001  -0.024  -0.030 
  7.413  -0.871  -0.147  -0.126  -0.030  -0.084  -0.079  -0.059   0.022  -0.044 
  7.564  -0.845  -0.074  -0.144  -0.059  -0.095  -0.087  -0.065   0.008  -0.023 
  7.521  -0.847  -0.164  -0.207  -0.023  -0.029  -0.068  -0.054   0.030  -0.019 
  7.582  -0.832  -0.114  -0.067  -0.049  -0.093  -0.066  -0.097  -0.070   0.005 
  7.638  -0.729  -0.162  -0.181  -0.134  -0.121  -0.022  -0.025  -0.022  -0.013 
  7.302  -0.685  -0.139  -0.247  -0.159  -0.084  -0.052  -0.047  -0.027  -0.012 
  6.462  -0.672  -0.304  -0.254  -0.010   0.064  -0.012  -0.147  -0.073  -0.002 
  6.334  -0.741  -0.349  -0.281  -0.011   0.051  -0.000  -0.195  -0.071  -0.019

Thanks!

[Do Quoc Truong]

I found another error:
ERROR:

"ngram_search.c", line 1144: Couldn't find <s> in first frame

[Nickolay V. Shmyrev]

The contents of the file look wrong. Most likely you are decoding the audio
with the wrong sample rate. The sample rate must be 8khz the feature
extraction configuration inside the tool must be updated accordingly.