Linear predictive coding

Atal and Hanauer described how to represent a speech signal by a dozen coefficients, predicting each sample from the preceding ones.

The method models the vocal tract as a filter and estimates its parameters instead of storing the waveform. Twelve coefficients replace hundreds of samples. That gave both compression for transmitting voice and features for recognition. The same technique spoke from the Speak and Spell toy seven years later and later underpinned mobile codecs.