4. Technical Issues

4.1 Sound Recording

Sound is recorded at 16 kHz, 16 bit, mono. This should be supported by any state-of-the-art standard sound card ?!

Note:

To be able to record sound data (on a Linux system) you need to introduce a user group that is allowed to access the sound device. In the following explanation I will assume that /dev/dspW is the sound device we want to use, audio is the name of the group that we want to grant access to the sound device and bill is the user id of the user who wants to use CVoiceControl.

To introduce the user group audio and to add the user bill to this group and to make sure that the new group may record from the sound device, open a console window and type (as root):

  adduser --group audio
  adduser bill audio
  chgrp audio /dev/dspW
  chmod g+r /dev/dspW

Now "ls -l /dev/dspW" should output something like

  crw-rw--w-    1 root     audio     14,   5 Jan  9 22:38 /dev/dspW

Then you have to restart Windows to complete the installation ;-)

For answers to further questions concerning Linux and sound have a look at the Linux Sound-HOWTO which is available from ftp://metalab.unc.edu/pub/Linux/docs/HOWTO/Sound-HOWTO

4.2 Recognizer Details

• Isolated word recognition
• Template matching using non-linear time normalization (DTW)
  • DTW, using symmetric Sahoe&Chiba warping function
  • Starting corner sloppiness: 4
  • Parallel warping planes and Branch&Bound recognition
  • Nearest neighbour wins template matching (roughly)
• Preprocessing:
  • Hamming window
  • Fast Hartley Transform (FFT) to calculate power spectrum
  • (log) melscale coefficients
  • Channel mean substraction
  • IMPORTANT: Amplitude normalization is not yet performed on the wave data so it is important to keep about the same distance between microphone and mouth and to speak at about a constant loudness level