I've trained a Sphinx2 acoustic model using 14 hours of speech data, and I've been testing it with a set of utterances that were held out of the training set. Results are OK -- good.
Then I noticed that the script I'd been using to run Sphinx (which was adapted from scripts/sphinx-train) had -agcmax TRUE, whereas I'd run the SphinxTrain scripts with the default CFG_AGC = 'none'. Oops! I've noted in several places in the Sphinx docs from the CMU website that it's stressed that the AGC parameter of the decoder must match that of the acoustic model. So I just made a new trial in which I changed only -agcmax to FALSE, and I was surprised to find that the matchfn file is IDENTICAL to before -- even the reported scores are the same. I did verify from the log files that AGC_MAX was running in the first case, with no AGC in the second.
So what could be happening to cause IDENTICAL results? Was the AGC-MAX actually doing nothing in the first case? Here are the sphinx parameters:
I've trained a Sphinx2 acoustic model using 14 hours of speech data, and I've been testing it with a set of utterances that were held out of the training set. Results are OK -- good.
Then I noticed that the script I'd been using to run Sphinx (which was adapted from scripts/sphinx-train) had -agcmax TRUE, whereas I'd run the SphinxTrain scripts with the default CFG_AGC = 'none'. Oops! I've noted in several places in the Sphinx docs from the CMU website that it's stressed that the AGC parameter of the decoder must match that of the acoustic model. So I just made a new trial in which I changed only -agcmax to FALSE, and I was surprised to find that the matchfn file is IDENTICAL to before -- even the reported scores are the same. I did verify from the log files that AGC_MAX was running in the first case, with no AGC in the second.
So what could be happening to cause IDENTICAL results? Was the AGC-MAX actually doing nothing in the first case? Here are the sphinx parameters:
sphinx2-continuous -verbose 9 \
-adcin FALSE \
-cepext feat \
-ctlfn ${CTLFILE} \
-ctloffset 0 -ctlcount 100000000 \
-datadir ${FEATS} \
-agcmax TRUE \
-langwt 6.5 \
-fwdflatlw 8.5 \
-rescorelw 9.5 \
-ugwt 0.5 \
-fillpen 1e-10 \
-silpen 0.005 \
-inspen 0.65 \
-top 1 \
-topsenfrm 3 \
-topsenthresh -70000 \
-beam 2e-06 \
-npbeam 2e-06 \
-lpbeam 2e-05 \
-lponlybeam 0.0005 \
-nwbeam 0.0005 \
-fwdflat FALSE \
-fwdflatbeam 1e-08 \
-fwdflatnwbeam 0.0003 \
-bestpath FALSE \
-kbdumpdir ${LMDIR} \
-lmfn ${LMDIR}/7568.lm \
-dictfn $youth/etc/youth.dic \
-noisedict $youth/etc/youth.filler \
-phnfn ${HMM}/phone -mapfn ${HMM}/map -hmmdir ${HMM} -hmmdirlist ${HMM} \
-8bsen TRUE -sendumpfn ${HMM}/sendump \
-cbdir ${HMM} \
-matchfn $youth/test/test.match_7000 \
-logfn $youth/test/test.log