-1

Recently I've focused on a project to implement a keyword spotting system. I've used HTK for speech recognition earlier. Now I want to know is it possible to implement my keyword spotter using HTK?

Termininja
  • 6,620
  • 12
  • 48
  • 49
Ehsan Maiqani
  • 1,684
  • 1
  • 14
  • 16

1 Answers1

0

Speech recognition and keyword spotting are quite related problems.

For HTK one of the two solutions is possible:

  1. build a word-loop grammar with a list of words you want to search, a garbage and a silence unit. See HBuild in HTKbook for details

  2. do a conventional speech decoding, which produces a word lattice (.slf in HTK). Then convert it in a consensus network (a sausage) with, for example lattice-tool, and search the words that have a score above some threshold

  • I have read the HBuild but find nothing abut my purpose. So I searched the web about how to create wordloop. But now I dont know what I should do. – Ehsan Maiqani May 06 '16 at 03:59
  • If your keyword list is not too large, you can build a network as described [here](http://www.ee.columbia.edu/ln/LabROSA/doc/HTKBook21/node133.html). You should also add in the network garbage and silence units. If using HTK is not obligatory, it can probably be easier for you to use [KWS from Sphinx](http://cmusphinx.sourceforge.net/wiki/tutorialandroid) – Arseniy Gorin May 06 '16 at 08:40