wave2feat / FeatureFileDumper same results?

[Hendrik Sirges]

Hi again. I have experimented a bit and tried to train an acoustic model. now im unsure if the features created from wave2feat and FeatureFileDumper are the same or which one is the "right one".
for example: i want to compute the mfc-file for a raw audio file.

my frontend.config.xml for the FeatureFileDumper is:
<component name="cepstraFrontEnd" type="edu.cmu.sphinx.frontend.FrontEnd">
<propertylist name="pipeline">
<item>streamDataSource </item>
<item>premphasizer </item>
<item>windower </item>
<item>fft </item>
<item>melFilterBank </item>
<item>dct </item>
</propertylist>
</component>

execute
ant -Dinput="test.raw" -Doutput="test.mfc" cepstra_producer

and execute
wave2feat -raw 1 -i test.raw -o test_w2f.mfc

As said in the Sphinx4FAQ the created files should be the same.

then if i look at the files with cepview:

cepview -d 13 -f test.mfc
it starts with:
14.150 -0.916 -0.246 -0.304 -0.160 -0.046 -0.099 -0.144 -0.085 -0.045 0.007 -0.232 -0.130

and cepview -d 13 -f test_w2f.mfc:
3.317 -0.770 -0.117 -0.176 -0.029 0.085 0.049 -0.016 0.044 0.074 0.139 -0.091 0.001

so not the same. i also tried other configurations in frontend.config.xml also not the same result.

So here my questions:
1. Does FeatureFileDumper expect raw audio input?

1. If SphinxTrain is setup to use "1s_c_d_dd"
   (i read this "1 stream 13cepstra 13delta 13double")
   features then why is the make_feats.pl script produces the features using wave2feat?
   (as said in Sphinx4FAQ wave2feat produces 13cepstra only without delta and double delta.)

greetings,

Hendrik

[Hendrik Sirges]

OK, I think i got it now.
SphinxTrain only wants the mel-features and if I
choose e.g. "1s_c_d_dd" it does not want the features in this format but it calculates the deltas and double deltas on its own.

is this correct?

[Aenima1891]

can someone answer to Sirges'question? i'm very interested about this topic. I have the same doubt..

please, answer..thanks