Ok - I admit it. The "voice recognition" is a hoax.
The voiceCommand class that runs on the NXT only counts sound pulses - the transition of sound sensor volume above a threshold and below the threshold.
The coding scheme is 1 pulse = left, 2 pulses = right, 3 pulses = back up and turn around. To fool the audience it is necessary to use "Left", "Go Right" and "Now turn backl"
Hand claps or words in any language work equally well.