Log on: Remember me
Powered by Elgg

Timo Baumann :: Blog Archives

May 2008

May 15, 2008

Klotz hin Gnubbel

This is what I get with the current acoustic model and a LM that was even trained including the correct sentence (und füge es ein in den Bauch des Elephanten).

Even using only just the correct sentence as a grammar returns und füge es,  instead of the complete sentence. The alignment shows, that es is supposedly spans the complete ein in den Bauch des.

I read that the current models are severely overtrained on one speaker, so I tried one of his utterances (de43-01, die Anwendung wird entwickelt) which is correctly understood if I use it as a grammar (effectively resulting in forced alignment) and which results in the beautiful phantasie wird entwickelt if I include this one sentence in the statistical LM as above. 

Thus, the bad results are probably due to the bad acoustic model. I've already uploaded the PentoNamingCorpus to Voxforge, thus hopefully, acoustic models will improve eventually. But if bad comes to worse, we'll have to train based on KCoRS and Verbmobil...

Keywords: ASR, Sphinx

Posted by Timo Baumann | 0 comment(s)