Speech recognition: Using julius and voxforge
First the grammar and vocabulary need compilation. To compile the mkdfa.pl script should be used. Note: On Ubuntu it is called mkdfa in stead of mkdfa.pl.
mkdfa.pl basics_and_shutdown
Run julius with the control script as
julius -C basics_and_shutdown.jconf | ./basics_and_shutdown.py
As you can see the control script is written in python v2.
Currently the syntax requires you to identify your computer as Lux (spoken as: looks), Lacer (spoken as: lazer), Alexia, or Pi.
After the identification follows a command. The following commands are possible:
Volume control uses amixer to set the Master volume.
- volume up/down
- volume 1,2,3,4,5,10,15,...,100
mute executes the commands amixer set Master toggle. Synonyms are:
- mute
- unmute
- shut up
- silence
The command script separates whether Alexia,Pi, or Lux,Lacer is used for identification. For Alexia and Pi, the commands correspond to the respective mpc commands. For Lux and Lacer, rhythmbox is used.
- next
- previous
- pause, unpause (spoken as: unpaw)
- stop
The following commands can be used to turn off, or prevent the previously commanded shutdown. Shutdown occurs in one minute.
- shut down (two separate words)
- cancel shut down (cancel with American pronunciation)
In order to allow for shutdown without password entry add the following line to your /etc/sudoers file. Replace user_name appropriately.
%user_name ALL=(ALL) NOPASSWD: /sbin/poweroff /sbin/reboot /sbin/shutdown
Currently we are unable to prevent julius from outputing the most likely command if no good match is found. Due to this reason the ID tag contains a number of ERROR outputs. This works for the moment.
The second big issue we encountered is due to to ambient noise level. Julius does not differentiate between sound output coming from the computer and external sounds. Using headphones there is no issue however with satellites louder volume can obstruct the required tag at the beginning and end each identification.