Fisher Notes

From SpeechWiki

accoustic features hain1999Htk talks about limiting the mel filterbank to telephone bandwidth 125hz=3800khz. performs worse than 0-4000khz if only mean normalizing, slight improvement if also var normalizing bigger improvement 1.3% if also gender dependent models.

rationalle for using _0 instead of _E for MVA is in http://ssli.ee.washington.edu/people/chiaping/mva.html but may be this is different for PLPs instead of MFCCs? but may be telephone is different from Aurora?...

many more spron/mpron datapoints in Hain 2005 Arabic dictionary paper ICASSP 2008 is possibly another data point