There were two tiers of synthetic speech in the 1980s, and which one you heard depended entirely on money.
The tier most people encountered was the SP0256 allophone chip, the part inside affordable home-computer speech add-ons. It stitched together phoneme fragments and produced the strangled, buzzing, unmistakably computer voice that the era’s films and games taught everyone to expect from machines.
The other tier was DECtalk — Digital Equipment Corporation’s professional synthesizer, expensive and genuinely intelligible, whose “Perfect Paul” voice became the most recognizable synthetic voice in the world when Stephen Hawking adopted it and then refused every later upgrade.
Michael Wessel’s Perfect Paul II, written up on September 14, 2026, closes that gap forty years late: a DECtalk emulator on a card that plugs into an Apple II.
How it works
The card carries an RP2040, an I2S DAC and amplifier, a speaker, and bus-interface logic to talk to the Apple II.
The design decision that makes it more than a curiosity is the interface. The card appears at a set of memory addresses, and driving it is, as the write-up puts it, “as simple as POKEing DECtalk commands to those addresses.” That’s the native idiom of the machine — POKE is how an Apple II programmer in 1983 would have expected to talk to any peripheral, and it’s two characters of BASIC away for anyone typing at the prompt today.
So this isn’t an emulator running beside vintage hardware. Existing 8-bit software can address it, and new BASIC written at the prompt can too, with no toolchain and no cross-compilation.
Why the anachronism is the point
A modern microcontroller pretending to be a 1983 expansion card is doing something more interesting than nostalgia. The RP2040 costs about a dollar and has capacity to spare emulating a synthesizer that once filled its own enclosure and cost more than the computer it was plugged into — which is a compact illustration of what forty years of silicon actually bought.
It also lands at a pointed moment. We wrote last week about sanoTTS, a neural text-to-speech model that fits in 337 KB and runs in real time on a $3 microcontroller. Perfect Paul II is roughly the same hardware class solving roughly the same problem from the opposite direction: one reconstructs a specific historical voice with reverence, the other synthesizes arbitrary new ones and doesn’t care what they sound like. Both now fit on a chip that costs less than a sandwich.
The project is open on GitHub.