| Abram Hindle's Blog | |||||
|
Subscribe
Flavours
Links |
Sun, 15 Jan 2006
Subject: Speech Recognition On Linux
I've tried different packages from CMU Sphinx2 Sphinx-2 Sphinx3 Sphinx-3 Sphinx4 Sphinx-4 I first tried Sphinx-4, it worked well for live demos but when I tried to integrate more complex languages into it, such as HUB4 or WSJ I just couldn't. Classpath problems and like even though I directly referenced both the jar files and class files. So Sphinx-4 was a bust simply do to it's very lame and over XMLify configuration system. Making if very difficult to get started. Then I started on Sphinx2, I tried to make my own LM files from random internet text and that was very troublesome. I could only use their web service to generate the files from text files. They only allowed for 5000 uniq words as well. Once I had that it was easy to run sphinx2-demo with the new ml and dic files (put into a new directory) except that it was very very picky of what I said. So picky that rarely did it ever print any output at all. Very disappointed I tried to figure out how to make it more liberal. I could not find the right parameters at all. So I was limited by vocabulary sizes, I was limited by the transitions between words in these vocabularies, and I was limited by the strict sensitivity of sphinx-2. I was rather unhappy. I had searched for general models etc and couldn't find them. Until I remembered, with Sphinx 4 I had tried to use HUB4 and WSJ data... So I tried to the HUB4 dump file with sphinx2 (to no avail). I then searched for hub4 and sphinx2... I found: http://www.arborius.net/~jphekman/sphinx/full/index.html So now I am waiting for this guarantuan file to load and see if it will recognize anything useful at all. Sadly it doesn't work well for general dictation at all :( abram |
||||