Print a Song on Paper with EnCodec
💡See how neural audio tokens shrink a 2.9 MB song to 21 KB—and why decoding and error recovery still matter.
⚡ 30-Second TL;DR
What Changed
EnCodec’s 3 kbps mode reduced a two-minute song to approximately 21 KB, shrinking the data by 99.9%.
Why It Matters
The project highlights how neural audio codecs can dramatically reduce storage requirements by representing sound as learned latent tokens rather than conventional compressed waveforms. It is more useful as an experiment in archival formats and ultra-low-bandwidth communication than as a practical music distribution method.
What To Do Next
Prototype EnCodec at 3 kbps on a short audio sample and measure reconstructed quality, file size, and decoder compatibility before designing an edge-audio pipeline.
Key Points
- •EnCodec’s 3 kbps mode reduced a two-minute song to approximately 21 KB, shrinking the data by 99.9%.
- •The encoded audio was divided across eight QR codes, with each code holding roughly 3.3 KB of binary data.
- •An ESP32 and RYLR998 LoRa setup could transmit the file reliably with acknowledgements and retransmissions, but the process took about 90 minutes.
🧠 Deep Insight
AI-generated analysis for this event.
🔑 Enhanced Key Takeaways
- •The experiment utilized the EnCodec model's ability to operate at extremely low bitrates by leveraging residual vector quantization (RVQ) to compress audio into discrete latent codes.
- •Makestreme's implementation relied on a custom Python script to convert the binary EnCodec output into a Base64-encoded format compatible with QR code density limits.
- •The use of QR codes for data storage introduces significant overhead due to error correction levels (ECC), which Makestreme had to balance against the physical size of the printed paper.
- •The LoRa transmission component utilized the RYLR998 module, which operates in the 868/915 MHz band, highlighting the trade-off between low-power long-range communication and low data throughput.
- •This project serves as a proof-of-concept for 'analog-digital hybrid' archiving, where physical paper acts as a cold storage medium for neural-compressed data.
🛠️ Technical Deep Dive
- EnCodec Architecture: Uses a convolutional encoder-decoder structure with a transformer-based language model for entropy coding, allowing for high-fidelity reconstruction at bitrates as low as 1.5 kbps to 6 kbps.
- Residual Vector Quantization (RVQ): The core mechanism that quantizes the latent representation into multiple codebooks, enabling the hierarchical compression used in this experiment.
- QR Code Constraints: Standard QR code (Version 40) capacity is limited to 2,953 bytes for binary data; splitting the 21 KB file required 8 distinct codes, each utilizing high-level error correction to mitigate printing artifacts.
- LoRa Protocol: The RYLR998 module was configured for high reliability, necessitating a custom packetization layer to handle the 90-minute transfer time due to the limited payload size per LoRa frame.
🔮 Future ImplicationsAI analysis grounded in cited sources
⏳ Timeline
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: IT之家 ↗