Speech remade from a buzz and a hiss
On 24 April 1939 Homer Dudley of Bell Telephone Laboratories read a paper on the automatic synthesis of speech before the US National Academy of Sciences; it appeared in PNAS that July. An analyser broke a talker's voice into 11 slow signals: one carried the pitch, ten the power in bands from 0 to 2950 cycles, each passed through a filter of up to 25 cycles. A synthesiser rebuilt speech from them out of two sources: a buzz from a relaxation oscillator for voiced sounds and a hiss from a gas tube for unvoiced ones.
Why it matters
A voice was split into two independent things, the source that excites it and the shape of its spectrum, and slow control signals proved enough to put speech back together. Dudley himself saw in this a way to private conversations and to a narrower band for telephone speech. Later parametric coding and synthesis of speech rest on that separation.
Pitch was measured by a channel of its own: a frequency discriminator, a frequency meter and a filter of 0 to 25 cycles; for unvoiced sounds its signal fell to zero and a relay switched the synthesiser from the buzz to the hiss. The first of the ten spectrum channels took 0-250 cycles and the other nine took bands 300 cycles wide up to 2950, the top chosen as representative of commercial telephone circuits; each held a band filter, a rectifier and a 25-cycle low-pass filter. The synthetic speech lagged the original by about 17 milliseconds. The analyser is shown at work in oscillograms of the sentence "She saw Mary", recorded on magnetic tape. Dudley describes the experimenter inverting inflections, holding a voice to a monotone, or forming a hybrid voice from the pitch of one talker and the spectrum of another. The October paper "Remaking Speech" in the Journal of the Acoustical Society adds, by its abstract, that the spectrum carries most of the intelligibility and pitch the emotional content. What the record does not claim: an intelligibility percentage or a total transmission band, because the paper gives neither. The word vocoder does not appear in it either.