NETtalk was an early text to speech application built using a multi-layer perceptron built in the mid 1980s.. In some way NETtalk's training is very smilar to more recent systems, using data form surrounding letters to predict the ecorrect phoneme to pronounce. However, in comparison the the many millions of neurons in current networks, NETtalk had a mere 203 input units, 80 hidden units and 26 output units. Yet despite this produced intelligable output and learning behavour rather like a child's. Certainly a lesson for those today trying to reduce the carbon footprint of AI and reduce the digital divide.
Used on page 114
Links:
Wikipedia: NETtalk (artificial neural network)