- Author
- John Vinyard
- Published
- Undated (the citation block gives the year 2025)
- Kind
- Research-style blog write-up of a proof-of-concept model, with audio examples
- Original
- blog.cochlea.xyz/sparse-interpretable-audio-codec-paper.html
- Subject
- Audio models and analysis
- Themes
- What the senses do, Art through code
What it is
The page is a research-style article by John Vinyard. It introduces a small model that splits audio into timed events. The stated goal is a representation people can interpret and compose in.
What it says
- The page says common codecs code fixed overlapping frames, which gives no readable representation.
- The page says the encoder removes energy step by step, producing events.
- The page says a 32-dimensional vector selects each event's attack, resonance and room.
- The page says masking half the input lets the model handle audio of any length.
- The page says the model was trained for about 76 hours on MusicNet.
- The page says the model has about 14M parameters, about 80 MB on disk.
- The page says reconstruction quality is not good enough for a lossy codec. It also says many events overlap.
What evidence it gives
The page offers five audio examples, each with the original and a reconstruction. It shows event scatterplots, intermediate decoder steps and randomized variants. No listening test or numeric quality score appears. Training code is linked.
Interests and limits
This is a proof of concept from one named author. The page itself says quality falls short of a usable codec. No comparison with other codecs is reported. Claims about future music models are speculation, which the page calls the authors' intuition. Only the page text was read. The audio and images were not listened to or viewed.
Connections
These links are this wiki's own reading, not claims the sources make.
- Synthesizing Music from JSON. Szablewski's songs are notes passed to instruments, usually fitting in 1200 URL characters. Vinyard's model tries to recover something similar from recordings: events with attack, resonance and room. One writes the score by hand. The other needs about 80 MB to guess it.