View Single Post
Posts: 959 | Thanked: 3,427 times | Joined on Apr 2012
#250
Originally Posted by myra View Post
One of the problems is, that pocketsphinx will always give you the best match it found, even when you said something (like "blahblahblah") that is not in your corpus/dictionary. pocketsphinx internally calculates a score but when using it with gstreamer there is no way (I know about) to get this value. I am a bit happy that I was able to patch the gst-plugin (gstpocketshinx) and to cross compile it for the N900. Since something like "blahblahblah" gives a significant lower score than a word which is in the corpus/dictionary this kind of "final results" can be sorted out.
This is a nice first step, but still I am trying to improve the accuracy. If anyone has managed to improve the recognition accuracy I would be really glad to get some help.

Greets
One thing that I hope for is runtime model/dictionary switching, to allow full dictation when sending e.g. a text. Hopefully I can use some gstreamer trickery (like using a filesink at the end of the pipeline), and re-try to understand what was said (with a larger dictionary/model) if the score is low...I hope there is some way to extract the score, if not I may have to dive into the plugin code and try to find a way to extract it.
 

The Following 4 Users Say Thank You to taixzo For This Useful Post: