Speak & Spell ships with the first single-chip LPC speech synthesiser

Filed under idea in canon/proposals.md, and idea is wrong. The proposal sits at line 278, inside the ## idea block that runs from line 195 to line 722, and the framing it offers — "a capability stops being called AI on the day it ships" — is a real pattern but it is not this object's to own. That framing is John McCarthy's complaint, it predates June 1978 by two decades, and this canon has already filed it: dijkstra-1959, itself an idea, carries it as the fourth of its four citation cases and calls itself "the reference specimen of the AI effect." An entry cannot descend from a framing and also be the framing. What is being filed here is a date on which a thing visibly happened in public, which is the definition eliza-1966 set when it argued its own way out of the moment kind, and which clippy-1996 and siri-2011 then reused: Dartmouth, Lighthill, Deep Blue, the LLaMA weights. In June 1978, at the Summer Consumer Electronics Show in Chicago, a machine that produced human speech from a model of the human vocal tract went on sale to the general public for $50. That had not happened before at that price or that scale, and everything about how the public has since expected machines to talk dates from the seven years that follow it.

Why not idea, stated at more length because the proposal file says otherwise. There is an idea inside this object, and it is a good one: pitch-excited linear predictive coding, in which speech is transmitted not as a waveform but as the time-varying coefficients of a filter that stands in for a throat and a mouth, driven by a buzz for voiced sounds and a hiss for unvoiced ones. But that idea is not TI's. It is Fumitada Itakura and Shuzo Saito's at NTT from 1966, and Bishnu Atal and Suzanne Hanauer's at Bell Labs from 1971, and in its ancestral form Homer Dudley's at Bell Labs from the 1930s. TI's contribution was not the idea; it was getting the idea to run in real time on 45,000 square mils of PMOS silicon that cost under five dollars to make. That is an engineering and manufacturing event with a date on it, which is a moment.

Why not limit. Nothing here is proved, and a good deal of what the object is used to prove is section 4's business. There is an interesting near-limit buried in it — the toy could not say a word that was not already in its ROM, and that constraint was economic and archival rather than theoretical — but a constraint you can buy your way out of is not a limit in the sense godel-incompleteness-1931 and turing-halting-1936 established for this canon.

On the date, which is layered and should not be smoothed over. Five dates matter here and they are not interchangeable. The idea was Paul Breedlove's in 1976, funded at $25,000 out of TI's internal idea programme; Richard Wiggins demonstrated the concept to Breedlove, Brantingham and Frantz in TI's Dallas research laboratory in December 1976. The patent's priority date is 17 June 1977 and its filing date 28 April 1978. The team had the product working in April 1978. It was shown at the Summer CES in June 1978 — the date the id carries and the one the header follows — and it reached shops in time for Christmas 1978. A reading that wants the date the thing existed should say April 1978; the date the trade saw it, June 1978; the date the public could buy it, autumn 1978. All three are correct and they are six months apart.

descends_from is empty, and that is a statement about the canon rather than about the object. Speak & Spell has at least four ancestors and none of them are written. The technical line runs Homer Dudley's Voder and vocoder (Bell Labs, demonstrated at the 1939 New York World's Fair) → Itakura and Saito's maximum likelihood formulation (NTT, 1966) and their PARCOR method (1969) → Atal and Hanauer, Speech analysis and synthesis by linear prediction of the speech wave (JASA vol. 50, 1971). The commercial line runs through TI's own TMS1000 calculator-on-a-chip and the Little Professor (1976), the product this was conceived as a successor to. The first-mover line runs through Forrest Mozer's algorithm, the TSI S14001A chip (1975) and the Speech+ talking calculator (1976) — see section 4, because that lineage is the one most often erased. None of these has an id. I have not invented any, per the job's instruction. If any one of them is later written, the strongest single edge to add is a Dudley entry, because the vocoder is the ancestor of the concept — analyse speech into a small set of control parameters, transmit those, resynthesise — and everything from LPC to the neural codecs of the 2020s is a refinement of it.

What it is

Speak & Spell is a hand-held orange plastic slab, about ten inches by seven, with a carrying handle, a vacuum-fluorescent display, forty raised keys in alphabetical order, and a speaker. It runs on four C cells. A child presses a button; the machine says, in a flat Texan baritone that sounds like it is talking through a mouthful of sand, "Spell… potato." The child types. The machine says "That is correct" or "Wrong. Try again," and after two failures spells the word out and pronounces it. The original had a little over two hundred commonly misspelled words for readers aged seven and up, extensible by ROM cartridges at about $20 each. It was manufactured from 1978 to the mid-1990s, spawned Speak & Read and Speak & Math in 1980, went to Tiger Electronics with TI's toy division in 1995 and to Hasbro in 1998, and was revived by Basic Fun in 2019.

The reason it is in a canon about artificial intelligence is one chip. TI's speech synthesiser — sold as the TMS5100, carrying the internal designations TMC0280 and, in some sources and on some dies, TMC0281 — was the first self-contained speech synthesiser on a single integrated circuit that worked by modelling the thing that makes speech rather than by storing and replaying the sound speech makes. Inside it is a ten-stage digital lattice filter. Ten reflection coefficients, plus an energy term and a pitch term, are pulled from ROM frame by frame and used to reconfigure that filter on the fly; the filter is driven by a stored glottal chirp waveform for voiced sounds and by a linear-feedback shift register generating noise for unvoiced ones. What comes out the other end is not a recording. It is a simulation of a throat, reconfigured about forty times a second, and it is why the voice has that specific buzzing-hissing-popping quality that no recording has ever had. The patent (US 4,209,836, Wiggins and Brantingham, granted 24 June 1980) describes clocking data out of the digital filter at a ten-kilohertz rate; the emulation and reverse-engineering literature that reconstructed the shipped part describes the lattice running at eight kilohertz. I could not reconcile those from the sources available in this run and record both rather than pick one.

The compression is the point. Two 128-kilobit metal-gate PMOS ROMs — 32 kilobytes in total, at a moment when 16 kilobits was the state of the art — held on the order of three hundred words and phrases. At roughly a thousand bits per second, speech that would have needed something like sixty-four kilobits per second as raw digitised audio fitted in a toy. The whole product is that ratio.

The voice belonged to a Dallas radio announcer, and the sources disagree about which one. Wikipedia and Adafruit's teardown both name Mitch Carr, an anchor at KRLD, chosen for a clear and level delivery. Richard Wiggins, commenting on his own project in 2015, named Hank Carr. Southern Methodist University's DeGolyer collection holds an item catalogued as "Voices for Speak & Spell. Jim Wiggins, George Doddington, Hank Carr, and Michael O'Shea," which suggests several people were recorded as candidates and that at least one Carr was among them. I could not open the SMU item (the CONTENTdm viewer returned no text to the fetch tool) and cannot resolve it. It is a small thing and it is exactly the sort of small thing a canon should not paper over.

The unglamorous half of the engineering was the corpus, and Wiggins is explicit about it. Analysing a human voice into LPC frames that a ten-pole filter could replay intelligibly was not automatic; frames had to be adjusted by hand. In his words, "hand editing was used… and wasn't always successful… consumed incredible amounts of time and was difficult to do." The consequence is section 2's most useful fact: Speak & Spell was not a text-to-speech system. It could not pronounce a word that had not been spoken by a specific man, analysed, hand-fixed and burned into mask ROM. Its vocabulary was not a capability. It was an inventory.

Why a reading would cite it

Five occasions, in descending order of how often they come up.

First, and much the strongest: when a demonstration's coverage is a curated list rather than a general capability, and nothing in the demonstration says so. This is the entry's real value and it is worth more than the AI-effect reading the proposal file offers. In 1978 a seven-year-old could hold in one hand a machine that spoke English. What that machine could actually do was replay two hundred words a man had said into a microphone in Dallas, compressed forty to one. The experience was "it can talk." The capability was "it has an inventory." Nothing in the object disclosed which — the toy did not, and could not, indicate that "aardvark" was outside its world in a different way than a misspelling was. A reading meets this shape constantly: a benchmark whose items were in the training set, an agent demo on the tasks it was tuned for, a coding model that is extraordinary on the languages with the most public code and mediocre elsewhere, a translation system with a hundred languages of which twelve are good. The question the entry supplies is precise and answerable: what is the machine's lexicon, and what happens at its edge? In 1978 the answer was knowable — you could count the words. That is what makes it a teaching case rather than an accusation, and a reading should use it to ask, not to conclude.

Second, when a capability is reclassified as ordinary the moment it gets cheap — with the caveat that this cuts both ways. Machine speech was a research problem in 1978. ARPA had spent roughly three million dollars a year for five years on the Speech Understanding Research programme at CMU, SRI, Lincoln Lab, SDC and BBN, wound up in 1976; lighthill-1973 already records that cancellation as the American half of the first winter, and records Lighthill's own 1973 judgement that money spent on recognising ordinary speech had been "wholly wasted." Two years after that programme closed, a $50 toy talked, and nobody called it artificial intelligence — not TI, not the trade press, not the parents. It was a toy. The entry supplies the deflationary question in the form dijkstra-1959 supplies it: when a 2026 product announcement leads with a synthetic voice, what is doing the work that was not available in 1978? The honest limit, which dijkstra-1959 states and which must be carried here too, is that the same reasoning deflates genuine achievements in retrospect, and this is how the field keeps losing credit for its wins. Also, and this matters: the two halves of the speech problem are not the same problem. ARPA funded understanding — getting meaning out of an acoustic signal. TI shipped generation. Speak & Spell did not solve what ARPA had failed at; it shipped the other side of the same subject, and any reading that uses it to score the SUR cancellation is comparing unlike things.

Third, as the dated floor of the synthetic-voice problem. Everything that is now a policy question about voices — cloning, consent, impersonation fraud, watermarking obligations for synthetic media, the whole 2026 apparatus — rests on the fact that a synthetic voice is now cheap, general, and indistinguishable from a particular named human's. Speak & Spell is the point where the first of those three arrived and the other two conspicuously did not. It was cheap in 1978. It was not general (see above) and it was not indistinguishable — its "technological accent" was so pronounced that it became the sound of machine for a generation and got sampled as such by Kraftwerk, borrowed for the title of Depeche Mode's 1981 debut, and adopted as an instrument by circuit benders who deliberately short its traces to make it stutter. A reading arguing about voice harms can date the gap: forty-eight years from the first mass-market synthetic voice to the first one you cannot tell from your mother's. What closed the gap was not one invention but the removal of two constraints — the inventory, and the accent.

Fourth, as a base rate for educational-technology claims, which is the one place this entry carries a graded outcome. TI sold it as a learning aid; Audrey Watters places it squarely in the teaching-machine tradition that runs through Pressey and Skinner, with immediate feedback and verbal reinforcement as the mechanism. A 1982 study in the International Journal of Man-Machine Studies found a significant increase in spelling of words in the machine's lexicon, that the increase decayed toward baseline once the machine was withdrawn, and that there was no improvement on words outside the lexicon. Learn the list, do not transfer, do not retain. That is a specific, dated, unflattering finding about the highest-selling educational electronic object of its era, and it is directly the shape of the claim being made in 2026 for AI tutors. It does not refute those claims — a 1982 result about a two-hundred-word ROM is not evidence about a 2026 tutoring system, and a reading that uses it that way is doing the thing section 4 warns against. What it supplies is the question to ask of any such claim, and the knowledge that the question has been asked before and the answer was disappointing: does the gain transfer off the list, and does it survive withdrawal? Source caveat, stated here rather than buried: I have this study only through Watters's account of it and could not retrieve its authors, title or volume in this run. It should be cited as "a 1982 study reported by Watters," not as a paper I have read.

Fifth, on the cost of data preparation as against the cost of the algorithm. Wiggins picked LPC in an afternoon of reasoning about what a chip could afford to compute. Getting a usable corpus out of it took the project's remaining time and was done by hand, frame by frame. The lesson generalises without strain to every era since — the imbalance between the elegance of a method and the brutality of preparing the data it eats is the most consistent unreported cost in this field, and norvig-data-2009 argues the same asymmetry from the other end.

What it got right, and what it got wrong

Not required for moment. This canon keeps doing it anyway, and there are four gradeable claims.

The patent's commercial claim. Made 1977–78, due immediately, correct beyond what it asked for. US 4,209,836 states as an object of the invention "to bring digital speech synthesis technology into a price range where the ordinary consumer may take advantage of it," and puts the manufacturing cost at "under five dollars when manufactured in quantity." The chip shipped in a $50 product that year — roughly $250 in 2026 money on the standard CPI adjustment — and the family it founded ran for nearly thirty years, through the TMS5110, 5200 and 5220, into Speak & Read, Speak & Math, the TI-99/4A speech module, the Apple II Echo card, the BBC Micro, and the arcade cabinets of Star Wars, Gauntlet and Paperboy. TI sold the line to Sensory Inc. in October 2001; Sensory stopped taking new mask submissions in October 2007. As a claim about price, this is about as clean a hit as the canon holds.

TI's educational claim. Made 1978, due within a few years, graded down. See section 2's fourth item. Gains on the lexicon, no transfer, no retention. TI was not lying — the machine did teach the words on it — but the claim as parents heard it was that the toy taught spelling, and that is not what was measured.

The engineering claim about the vocal tract, which is right and is routinely overstated by one step. The line that recurs in the retrospectives is that this was "the first time the human vocal tract had been electronically duplicated on a single chip of silicon," and as stated that is fair: a ten-pole lattice filter with a glottal source is a model of a vocal tract, and it was on one chip. The overstatement, which the same retrospectives then make, is to describe the toy as generating speech from phonemes or from text. It did neither. It replayed analysed recordings of one man through that model. The distinction is the whole of section 2's first item, and losing it is the most common error about this object.

The IEEE's claim, made in 2009, about what began here. The IEEE named it a Milestone in Electrical Engineering and Computing in 2009 — "Speak & Spell, the First Use of a Digital Signal Processing IC for Speech Generation, 1978" — with the plaque at TI's North Campus on North Central Expressway in Dallas, and the supporting case describes the TMS5100 as the earliest digital signal processing chip in the industry and the start of the DSP business. I could not read the plaque citation verbatim: ethw.org returned HTTP 403 to every fetch attempt in this run, and what is above is assembled from search-result extracts of that page. Treat the wording as approximate and the substance as sound. The claim that this specific chip founded DSP is a Milestone-nomination claim, made by interested parties about their own employer's product, and a reading should say so if it leans on it.

Commonly misused as

Not required for moment. It earns a section here because the popular account of this object is wrong in four specific ways, three of which are firsts it did not achieve.

• "The first talking machine." No. Homer Dudley's Voder produced continuous synthetic speech electronically at the 1939 New York World's Fair and the Golden Gate International Exposition, played like an organ by around twenty operators trained by Helen Harper — a keyboard for the bandpass gains, a foot pedal for pitch, a wrist bar to switch buzz for hiss. Thirty-nine years earlier, in public, to enormous crowds. What Speak & Spell added was that no trained operator was required.
• "The first speech synthesis chip," or "the first talking consumer product." Also no, and this is the erasure worth correcting because it has a named victim. The TSI S14001A, designed by Ed Bernard at Silicon Systems to implement Forrest Mozer's algorithm, was a single-IC speech synthesiser in 1975, and it shipped in TeleSensory Systems' Speech+ hand-held talking calculator in 1976 at $395, built for blind users, with a 24-word vocabulary and its manual on cassette. Two years before Speak & Spell, in a consumer's hand. The correct and still-substantial claim is narrower: the TMS5100 was the first single-chip synthesiser that worked by modelling the vocal tract — the S14001A stored delta-modulation-compressed waveform samples — and Speak & Spell was the first time synthetic speech was sold as a mass-market object at a toy price rather than as assistive equipment at professional prices. That the assistive product came first, and is the one nobody remembers, is itself worth a reading's attention when the subject is who gets credited for a capability.
• "It did text-to-speech / phoneme synthesis." No, and this error is printed in otherwise careful retrospectives. Fixed inventory, mask ROM, hand-edited frames, one man's voice. The capability to speak arbitrary text arrived later in the same chip family and was a separate achievement.
• "It proves AI capabilities always commoditise into toys." A real pattern used as a law. It commoditised because the specific thing being shipped was replay of a fixed list, which is a bounded problem that fits in mask ROM, and because TI already owned a calculator-chip production line to put it on. The general problem — speaking anything, in anyone's voice — took another four decades and is a different kind of thing. Using 1978 to predict that any given 2026 capability is about to become a $50 object requires showing that the capability in question is bounded the way this one was, and that argument is almost never made.

Sources

A caveat that applies to everything below, in the form deep-blue-1997 established for this canon and for the same reason. Almost none of this material was read by me directly. The fetch tool available in this run returns another model's summary of a page, so quotations arrive at one remove and could have been introduced in summarisation. Where I quote, I quote what came back; where two sources disagree I have said so rather than picked. Direct-fetch failures in this run, recorded so the next session does not repeat them: ethw.org and ieeemilestones.ethw.org returned HTTP 403 (so the IEEE Milestone citation text is from search extracts only); datamath.org failed TLS negotiation and is unreachable from here despite being the best hobbyist source on TI chip part numbers; direct.mit.edu returned 403 on the open-access monograph chapter "4 SPEAK & SPELL (1978)", which is the one scholarly treatment I found and could not read; ieeexplore.ieee.org returned an empty body for Frantz and Wiggins's own 1982 account, "The development of 'Solid state speech' technology at Texas Instruments"; SMU DeGolyer's CONTENTdm viewer returned no text for the voice-recordings item. Outbound network access from the shell was not permitted, so no fallback fetch was possible.

Primary.

• US Patent 4,209,836, Speech synthesis integrated circuit device, Richard H. Wiggins Jr. and George L. Brantingham, assigned to Texas Instruments. Priority 17 June 1977, filed 28 April 1978, granted 24 June 1980. Read via Google Patents. Source of the ten-coefficient lattice, the ten-kilohertz filter output rate, the 45,000-square-mil active area, the under-five-dollars manufacturing target, and the quoted object of the invention. Note that secondary listings (Justia) give a filing of 19 June 1978 and a grant of 18 November 1980; those dates do not match the patent record as returned and are probably a different member of the family.
• Richard Wiggins, interviewed by Benj Edwards, VC&G Interview: 30 Years Later, Richard Wiggins Talks Speak & Spell Development, vintagecomputing.com. First-person, and the source for why LPC was chosen, for the hand-editing quotation, for the confirmation that arbitrary words were out of reach, for the $50 price and the Summer CES debut, and for the "Hank Carr" attribution.
• IEEE Milestone, Speak & Spell, the First Use of a Digital Signal Processing IC for Speech Generation, 1978, dedicated 2009, plaque at TI North Campus, 13532 North Central Expressway, Dallas. Via search extracts only — see the caveat above.

Technical.

• Wikipedia, Texas Instruments LPC Speech Chips — the chip family from TMS5100 to the TSP50Cxx, the pitch-excited LPC description, the glottal-pulse and LFSR excitation, the product list, the 2001 sale to Sensory and the 2007 end.
• Wikipedia, Speak & Spell (toy) — tenth-order LPC, ~1000 bits per second, the 128-kilobit PMOS ROMs, the TMS1000, the 200-word vocabulary, the "Mitch Carr, KRLD" attribution, the 1992 end of the original line and the 2019 revival.
• Adafruit Learning System, Bringing Back THE VOICE of Speak & Spell, Phillip Burgess — the 32-kilobyte ROM and ~300 words and phrases, the LPC-10 lineage, and the observation about the "technological accent" that LPC imposes on one specific person's voice.
• Mutable Instruments technical notes and the ti_lpc decoder project — the ten-stage lattice at 8 kHz, which is the figure that conflicts with the patent's 10 kHz.

Historical and contextual.

• IEEE Spectrum, Chip Hall of Fame: Texas Instruments TMC0281 Speech Synthesizer, and The Consumer Electronics Hall of Fame: Texas Instruments' Speak & Spell — the four-person team, the $25,000 idea-programme funding, the pair of 128-kilobit ROMs against a 16-kilobit state of the art, Frantz's "it was impossible to do in an integrated circuit," and the note that parents were dismissive in test marketing.
• Tedium, Speak & Spell History: Texas Instruments' Greatest Product (19 June 2018) — the Little Professor lineage, the $2 TMS1000, the portability constraint that ruled out contemporary high-performance methods, the $50 price.
• Audrey Watters, Speak & Spell: A History — Toys as Teaching Machines, hackeducation.com, 13 January 2015 — the Skinner/Pressey framing, TI's marketing language, and the 1982 International Journal of Man-Machine Studies finding. The study itself is unverified at first hand; see section 2.
• On the ancestry, none of which is in canon/: Wikipedia's Voder and Homer Dudley (the 1939 World's Fair demonstration, the twenty operators, Helen Harper); Wikipedia's Fumitada Itakura and the linear-prediction history literature (Itakura and Saito 1966 and PARCOR 1969; Atal and Hanauer, JASA vol. 50 pp. 637–655, August 1971).
• On the prior consumer product: Computer History Museum's catalogue entry for the TSI Speech+ model S1C (24-word vocabulary, manual on cassette, designed for the blind), vintagecalculators.com's Speech+ and Development of the TSI Speech+ pages (the $395 price, 1976 on sale, Telesensory of Palo Alto), and the S14001A designer's account via the chip's documentation (Ed Bernard at Silicon Systems, Forrest Mozer's algorithm, floating-zero two-bit delta modulation).
• On the ARPA context, and cross-referenced against this canon's own lighthill-1973: the SUR programme 1971–1976, Newell's 1971 study group, the thousand-word target, roughly $3m a year for five years, Harpy's 1,011-word vocabulary and the September 1976 demonstrations.
• On the cultural afterlife: Wikipedia's Speak & Spell (album) for Depeche Mode's 1981 debut title, and the Computer World (1981) usage by Kraftwerk.