Nearly every machine here is a source and a filter: a buzz (the voiced source, the vocal folds) or a hiss (the unvoiced source, turbulence) shaped by resonances (the vocal tract's formants). What changes from machine to machine is how the filter is made and who drives it:
Wins the Russian Imperial Academy's prize with resonators that produce the five long vowels. W
Publishes his speaking machine, built over twenty years: bellows for lungs, a reed for the voice, a leather tube the operator shapes by hand. W, L
Builds his own version of von Kempelen's machine; it makes vowels and most consonants. L
Shows the Euphonia, a talking head with a keyboard. S
Vowel sirens (sirènes à voyelles): plaster casts of real mouths, lips and teeth included, each shaped for one vowel, sounded by a siren. He used them to measure how far each vowel carries. The subject of Revue Faire #47/48. F
The Voder at the New York World's Fair: a buzz and a hiss through ten band filters, played by a trained operator on keys, a wrist bar and a foot pedal for pitch. The same lab's vocoder split a voice into those bands to send it in less bandwidth. W, S
Visible Speech (Van Nostrand): the sound spectrograph, frequency against time with darkness for energy, built so deaf people could read speech. Every spectrogram since descends from it, and Pattern Playback ran it backwards. N
Pattern Playback turns a painted spectrogram back into sound: you could paint speech and hear it. Sources disagree on the year. L
The first formant synthesizers: a few resonances, each set by a number. Details not yet checked against Klatt 1987. L
An IBM computer sings "Daisy Bell", Mathews providing the accompaniment. Arthur C. Clarke heard it, and HAL sings it in 2001. Sources name the machine as the IBM 704, 7090 or 7094. W
The Speak & Spell, released June 1978, built on the first single-chip speech synthesizer (TMS5100, internally TMC0280): linear predictive coding, a recorded voice compressed into a small digital filter and its excitation. Wiggins brought LPC; Breedlove had the idea of a talking toy; Brantingham laid out the silicon. An IEEE Milestone. M, I
The SC-01: 64 phonemes from a formant model, chained by a computer, with two bits of inflection. Heard in arcade games (Gorf, Q*bert). Made until at least 1984. V
Publishes a cascade/parallel formant synthesizer in the Journal of the Acoustical Society of America; it becomes the reference model. C
The SP0256-AL2 "Narrator": about 59 allophones in ROM, sold by Radio Shack; a sibling mask, the SP0256-012, holds the Intellivoice's phrases. Exact release year unchecked. P
S.A.M., Software Automatic Mouth: speech with no chip at all, on the Atari 8-bit, Apple II and Commodore 64; about 50 phonemes in an ASCII notation. Later the basis of the first MacinTalk. A
DECtalk, Klatt's synthesizer as a product. D
Every machine in the table makes a source and filters it. The mechanical ones did it with reeds and leather; the Voder with ten band filters and a pair of hands; the formant machines with a few numbers per sound; TI's chip with a filter fitted to a real voice; S.A.M. with nothing but a program. The operator never went away either: von Kempelen's hand on the tube, the Voder's trained player, and the Speak & Spell's child pressing letters are the same job at different distances.
Rebuilt in a few lines of numpy each: a buzz or a hiss and some filters, nothing recorded.
One 120 Hz buzz through three resonators set to "ee", "ah" and "oo". Only the filter changes.
Ten band filters on a buzz, the levels gliding from "ah" to "ee" to "oo" while the pitch bends, then the hiss with the high bands up for an "s".
Three painted spectrograms, the same vowel, with the second formant starting at 720, 1800 and 3000 Hz: the loci Delattre, Liberman and Cooper found for b, d and g.
| name | maker, year | method | heard in | can you get one? | emulated by |
|---|---|---|---|---|---|
| TMS5100 (TMC0280) | Texas Instruments, 1978 | LPC: a 10-pole digital filter and its excitation, fitted to a recorded voice; fixed vocabulary in ROM | Speak & Spell | in old Speak & Spells; the toy is the usual source | the Talkie Arduino library plays TI LPC data (unchecked); MAME (unchecked) |
| TMS5200 / TMS5220 | Texas Instruments | LPC, as above | the TI-99/4A Speech Synthesizer; arcade games | used parts turn up (unchecked) | Chipspeech's "Terminal 99" (from TI ROM data) |
| Votrax SC-01 / SC-01A | Votrax (Federal Screw Works), 1980 | formant model, 64 phonemes and 2 inflection bits, chained by the host computer | Gorf, Q*bert, talking terminals | rare | Chipspeech's "Bert Gotrax" (from recordings of an SC-01) |
| SSI-263 | Silicon Systems | phonemes, formant | the Mockingboard card for the Apple II (unchecked) | rare (unchecked) | |
| SP0256-AL2 "Narrator" | General Instrument, late 70s or early 80s | about 59 allophones in ROM driving a 12-pole LPC filter, six two-pole sections, at 10 kHz, per MAME's emulation | Radio Shack's speech kits (the -012 mask, with Mattel's phrases, is in the Intellivoice) | discontinued; still on eBay, but many sold online are fake; Arduino and RC2014 boards exist | Chipspeech's "Spencer AL2" (from Microchip ROM data) |
| Digitalker (MM54104) | National Semiconductor | compressed recorded words, fixed vocabulary (unchecked) | |||
| S.A.M. | Mark Barton, Don't Ask Software, 1982 | software only, about 50 phonemes in an ASCII notation | Atari 8-bit, Apple II, Commodore 64; the first MacinTalk | free ports exist (an Atari 2600 port in 2020) | many software ports |
| DECtalk | Digital Equipment Corporation, 1984 | Klatt's formant synthesizer | talking terminals, telephone services | used units | Chipspeech has a DECtalk-style voice |