Python HTK Converter

If you are using HTK still and has headaches with it, you have last chance to move to CMUSphinx! You can even take your models with you!

The htk2s3conv package which converts HTK model to SphinxTrain format has been landed in our VCS recently, check it out at

https://cmusphinx.svn.sourceforge.net/svnroot/cmusphinx/trunk/htk2s3conv

Don't forget to read the documentation!

Sphinx4 Creates Future

Recently we wrote about one interesting project where sphinx4 was used in contemporary art installation. That's really enouraging that being very easy for art designer sphinx4 allows to create new virtual reality. Please consider another interesting one:

Chatter from Mia Sorensen on Vimeo.

If you got interested, code and description is here.

Pocketsphinx Is A Perfect Fit For Mobile

Pocketsphinx is going to win on mobile platforms. Right now there is no alternative to using it because of performance, accuracy and flexibility. Using pretrained US English acoustic models one can get amazing results.

That was proved once again by amazing Nightingale Browser Project. Check this video how it works

If you like it, feel free to contribute to the project. Checkout source from Gitorious http://gitorious.org/nightingale, clone, send patches, report bugs!

And, to not make you confused. Pocketsphinx is not only for mobile. You can perfectly use it in desktop applications, even in large scalable application. It's accuracy is the best we can suggest you right now!

PocketSphinx 0.6.1 release

This is a bugfix release, addressing a number of important issues in the 0.6 release. Specifically:

  • The GStreamer plugin was broken with old versions of GStreamer (such as the one shipped on Nokia Internet Tablets) which did not accept "BSD" as a valid license type.
  • Runtime performance of the statistical LM based decoder was significantly worse than 0.5 when not using phoneme lookahead.  It is now about 10% faster and also uses less memory.
  • The FSG decoder now consumes drastically less memory and CPU power.  (It is still not as good as it should be, but no longer completely embarrassing)
  • The Python modules were undocumented.  Docstrings compatible with epydoc have been added.
  • Raw audio logging (-rawlogdir) was broken.
  • Adding new words works properly, including words with unknown triphones (which never worked before).
  • The SphinxBase Python module can now use class-based language models.
  • The pitch estimation utility (sphinx_pitch) now builds properly on Win32.
  • Some problems with the N-Gram iterator API have been fixed.
  • Bestpath search has been fixed to handle recognition failure gracefully.

Source code is available for download on SourceForge.  Binary packages for Ubuntu will be updated shortly.