Song on Paper: How an AI Codec Compressed a Track 1000x and Encoded It in QR Codes

Engineer Makestreme presented a project where a two-minute composition was compressed from 2.9 MB to just 21 KB and printed on paper as eight QR codes. This was achieved using Meta's neural audio codec EnCodec, released in 2022. The codec converts audio signals into tokens, which are then decoded back into sound. Settings at 3 Kbps preserved acceptable quality, whereas 1.5 Kbps introduced significant distortion.

Each QR code holds about 3 KB of data, requiring eight codes—arranged on both sides of a sheet like a vinyl record—to store the entire song. To maximize data density, error correction was minimized, making the recording highly vulnerable to damage: even a single corrupted code could render the composition unplayable.

The creator also demonstrated transmitting the compressed file via a LoRa radio channel between two ESP32 boards. Due to conservative protocol settings, the transfer took nearly 90 minutes. For comparison, another developer recently achieved speeds of 190 KB/s by transmitting QR codes between smartphones at 60 frames per second.

The project echoes vintage data storage methods, such as the 1985 Cauzin Softstrip, which printed programs as barcodes in magazines. However, the modern approach relies on neural networks, making the data dependent on the decoder's availability. While paper can last centuries, the song remains accessible only as long as the neural network capable of decoding it exists.