The Premise
Take the newest thing (AI music synthesis, generative audio) and put it on the oldest thing in home-computer storage (the Commodore Datasette cassette tape). Ask the AI to do two jobs at once: write a piece of music that listens like music, and embed a hidden payload in the spectral content that a different kind of listener can extract.
Think Aphex Twin's Windowlicker trick (the spectrogram face hidden in the noise toward the end of the track) but on purpose, at scale, with AI doing both halves of the job.
Then store it on a tape that played on a Commodore 64 in 1982.
Why Tape Now
The Commodore Datasette (technically the C2N or 1530) was the dominant home-computer storage medium between the late 1970s and the mid-1980s. It encoded data as audio tones at roughly 300 baud, stored each bit as a frequency pulse, and used a redundant double-write scheme so reads could survive the tape stretch. A typical program took five to ten minutes to load. The screech during load is iconic.
Cassette is still readable. Drives still exist on eBay. The encoding format is documented. Brand-new audio mastered to cassette today plays back on a 1982 machine. That's a fact worth using.
Tape predates floppy in the home-computer timeline. Floppies were invented in 1971 (IBM) but didn't displace tape for home use until the late 80s. The Commodore 64 launched in 1982 with the Datasette as the default storage. So a tape played on that machine sits at a real historical pivot.
The Aphex Twin Move
Two famous examples of audio steganography:
- Windowlicker (1999), Aphex Twin. The B-side track Δ Mᵢ⁻¹ = -α Σ Dᵢ[n] etc. ends with a spectrogram of Richard D. James's face. You can't hear the face. Run it through a spectrogram visualizer and there it is.
- [Equation] and Looking Glass (Plaid, Boards of Canada, others) embed similar steganographic objects. Some are decorative. Some are hostile. Most are private jokes.
What makes the move work is the hidden layer doesn't degrade the listenable layer. The music stands on its own. The hidden image only matters if you go looking. AI is good at this kind of dual optimization: write a tune that scans as music to ears, scans as a payload to spectral analysis, and pass both tests cleanly.
What's the Payload
This is where the project gets fun. Open question, maybe permanent, maybe each tape gets a different one:
- An ASCII portrait of someone the listener doesn't know. They find a face in the noise. They never know whose face. The face's identity stays on the tape and in my head.
- A short message in plaintext. A poem. A piece of advice. A confession. A coordinate. A name.
- A cryptographic key. One-time pad, public key, signed manifest. The tape becomes a physical artifact for digital trust.
- A SETI-style cosmic message. The Arecibo broadcast was 1,679 bits. Plenty of room on a cassette. Encode a small message intended for whoever finds the tape after enough time has passed. The tape becomes a deliberate artifact-for-future, easier to recover than a satellite.
- A photograph. Audio-encoded image, recoverable by reversing the encoder. Small, lossy, real.
- An entire small program. A Commodore 64 BASIC program that runs when loaded. Not a payload to be revealed via spectrogram, but the original Datasette use case. Audio + payload IS the format. The hidden bit is just whether the audio is also musical.
Different tapes can carry different payloads. Some tapes can carry layered payloads. Some tapes can be released publicly (with no hint of the hidden content) and see who notices.
How AI Plays Both Sides
The interesting work is the dual optimization:
- AI generates a piece of music. Style of the user's choice. Constraints on key, tempo, instrumentation.
- The hidden payload (image, text, key, program) is encoded as a target spectral pattern.
- The model adjusts the music's frequency content so the hidden spectral pattern is present but masked under the listenable layer.
- Output is a single audio file that listens normally, and yields the hidden payload when run through a spectral analyzer or the encoded extractor.
- That audio gets mastered to cassette via real tape (not just a digital file pretending). The cassette gets played on a Commodore 64. The hidden content survives the tape's analog noise floor because the encoding leaves headroom.
Getting the encoding to survive analog tape generation is half the technical work. The other half is the music holding up as music.
Could It Be A Record
Yes. Same idea on vinyl. The vinyl press introduces its own analog noise but spectral steganography survives a clean cut. A 12-inch single with a hidden payload nobody is looking for would be a beautiful object.
The tape is the v1. The vinyl is the v2.
Would Anyone Notice
That's part of the experiment. A piece of music released publicly with no claim of hidden content, that turns out to be a Trojan for a real message: if nobody decodes it, the hidden message is preserved by being unrecognized. If someone does decode it, the project becomes about the act of looking.
Both outcomes are interesting.
Status
Concept. Hardware on hand or trivially acquirable: Datasette drives are on eBay for under a hundred dollars, working Commodore 64s for two hundred. The encoding pipeline (audio steganography library + AI music generator + cassette-write utility) is the build.
Open questions worth real research: which audio-stego encoder survives cassette transfer best, how much payload bandwidth is realistic at 300 baud equivalent, whether to use the Datasette's native encoding format (so the tape actually loads as a Commodore program) or a modern stego scheme on top of conventionally-mastered audio.
The thing I am most sure of: the right answer is to ship a real, listenable, beautiful piece of audio that nobody knows is also data, and let the world figure it out (or not). Side-project tier.